You are on page 1of 12

Meteorol. Atmos. Phys.

71, 255±266 (1999)

1
Regional Meteorological Centre, Civil Aerodrome, Guwahati, India
2
India Meteorological Department, Pune-5, India

Long Range Forecasting Indian Summer Monsoon Rainfall


by a Hybrid Principal Component Neural Network Model
P. Guhathakurta1 , M. Rajeevan2 , and V. Thapliyal2

With 8 Figures

Received August 20, 1998


Revised April 20, 1999

Summary range forecasts of summer monsoon rainfall over India as a


whole.
The existing methods based on statistical techniques for
long range forecasts of Indian monsoon rainfall have shown
reasonably accurate performance, for last 11 years. Because 1. Introduction
of the limitation of such statistical techniques, new
techniques may have to be tried to obtain better results. In Long range forecasts of Indian summer monsoon
this paper, we discuss the results of an arti®cial neural rainfall (June±September) have been very crucial
network model by combining two different neural networks, for proper agricultural planning in India. Besides,
one explaining assumed deterministic dynamics within the the summer monsoon rainfall activity over India
time series of Indian monsoon rainfall (Model I) and other
using eight regional and global predictors (Model II). The
also has considerable impact on other national
model I has been developed by using the data of past 50 activities such as power generation, water resources
years (1901±50) and the data for recent period (1951±97) etc. After the initial work of Walker (1910), several
has been used for veri®cation. The model II has been attempts (Thapliyal, 1982; Gowariker et al., 1989,
developed by using the 30 year (1958±87) data and the 1991) have been made for developing better
veri®cation of this model has been carried out using the
models of long range forecasts of summer mon-
independent data of 10 year period (1988±97). In model II,
instead of using eight parameters directly as inputs, we have soon rainfall of India. Performance of the Param-
carried out Principal Component Analysis (PCA) of the etric and Power Regression models (Gowariker
eight parameters with 30 years of data, 1958±87, and the et al., 1991) have been satisfactory and reason-
®rst ®ve principal components are included as input ably accurate during last 11 years. These models
parameters. By combining model I and model II, a hybrid are used by the IMD for long range forecasts of
principal component neural network model (Model III) has
been developed by using 30 year (1958±87) data as training
summer monsoon (June±September) rainfall over
period and recent 10 year period (1988±97) as veri®cation India as a whole. Since these statistical models
period. Performance of the hybrid model (Model III) has have some inherent limitations, emphasize is given
been found the best among all three models developed. to develop better and alternate techniques for
Root mean square error (RMSE) of this hybrid model during long range forecasts of summer monsoon rainfall
the independent period (1988±97) is 4.93% as against
of India.
6.83% of the operational forecasts of the India Meteor-
ological Department (IMD) using the 16 parameter Power In this paper, an attempt is made to explore the
Regression model. As this hybrid model is showing good potential of neural network technique for long
results, it is now used by the IMD for experimental long- range forecasts of Indian summer monsoon
256 P. Guhathakurta et al.

rainfall. Recently this technique has drawn model based on neural network technique
considerable attention from research workers performed better.
as it can handle the complex and nonlinear We have used the neural network technique to
problems better than the conventional statistical develop three different types of models for long
techniques and has successfully been applied to a range prediction of summer monsoon rainfall
variety of problems. Neural network technique over India. In the ®rst, we have used only time
has a strong potential for pattern recognition and series data of monsoon rainfall as input to predict
signal processing problems and it has the ability the monsoon rainfall for the future. This model
to predict for the future value of the time series. (Model I) reconstructs the assumed deterministic
This technique has successfully been applied to a dynamics of the monsoon rainfall data. We have
variety of problems. It has been shown by Elsner further used eight regional and global parameters
and Tsonis (1992) that neural network can be as the predictors which are chosen by examining
used successfully to predict a chaotic time series. their physical linkage with the monsoon and their
Application of neural network for short term degree of relationship with the monsoon rainfall
prediction of air pollutants (Bonzar et al., 1993; of India. Since some of these predictors are inter-
Guhathakurta, 1998) has shown interesting correlated, we have carried out principal com-
results. The neural network technique is also ponent analysis of these eight parameters and
able to learn the dynamics within the time series the ®rst ®ve principal components are taken as
data (Elsner and Tsonis, 1992) and long range inputs to develop a principle component neural
prediction models of monsoon rainfall of India network model (Model II). Finally these two
using only time series data of monsoon rainfall models are combined by a two layer (without
have shown some encouraging results (Goswami hidden layer) hybrid neural network model
and Srividya, 1994; Guhathakurta and Thapliyal, (Model III). The performance of the hybrid
1997). However the model using only time series model (Model III) has been greatly improved
data of rainfall alone can not give reasonably over other two models.
accurate forecasts. The accuracy of the model This paper is arranged as follows. In Sect. 2,
using time series of rainfall can be improved by details of the data used are given. In Sect. 3 the
including some of the signals from the coupled neural network techniques is reviewed. In Sect. 4
surface-atmosphere system, as the inter-annual to 7 results are discussed and in Sect. 8 con-
variability of Indian summer monsoon rainfall is clusions are drawn.
in¯uenced by slowly varying boundary condi-
tions like sea surface temperatures (SST), and
2. Data
snow cover (Shukla, 1987). Recently, Venkatesan
et al. (1997) have used neural network technique All India summer monsoon rainfall as a percen-
to predict monsoon rainfall of India by using few tage of long period average is used in this study.
predictors and compared the results with linear For 96 years (1901±1996), the monsoon rainfall
regression techniques. They have found that the data are taken from available literature (Tha-

Table 1. Details of the 8 Predictors Used in this Study


S. No Parameter C. C. 1958±97 Reference
1. Darwin Pressure Tendency (April±January), DPT ÿ0.54 Shukla and Mooley (1987)
2. East Coast of India Minimum Temperature (March), ECT 0.52 Gowariker et al. (1991)
3. Nino 3 Index Tendency (MAM-DJF), NI3l ÿ0.36 Krishna Kumar et al. (1995)
4. NW Europe Mean Temperature (January), NWET 0.42 Rajeevan et al. (1998)
5. NW India Pressure Anomaly (MAM), NWPA ÿ0.52 Krishna Kumar et al. (1995)
6. NH Pressure Anomaly (Jan±April), NHPA 0.55 Gowariker et al. (1991)
7. 10 hPa Zonal Wind Anomaly at Balboa, 10HZA 0.42 Bhalme et al. (1987)
8. Minimum Temperature of NW and Central India (May), 0.41 Krishna Kumar et al. (1995)
MAYTM
Long Range Forecasting Indian Summer Monsoon Rainfall 257

pliyal, 1997). The series has been updated up to ruggedness, (iv) speed (via massive parallelism),
1997. This time series has been prepared as area (v) nonlinearity and (vi) optimality with respect to
weighted mean of seasonal (June±September) error. For regression, we assume a functional form
rainfall of 35 meteorological sub-divisions of ®rst, such as linear or exponential, and then we
India by making use of rainfall data of more than ®nd the coef®cients that minimize some measure
4000 stations. This is the time series used by the of errors, whereas for neural networks, the method
IMD for operational long range forecasting. itself extract the functional form from the data. As
The eight parameters used as the predictors are input to the model, a historical set of signi®cant
given in Table 1. Data of these eight parameters meteorological data is used, where as the output,
for the period 1958±1997 were collected from monsoon rainfall is predicted by the model. The
``Monthly climate data for the world'', and network is trained with the past data. By the
monthly climate records of the IMD. Nino-3 proper choice of training sets, after the learning
index was obtained from the Climate Analysis process, the trained network is capable of
Centre, National Centre for Environmental Pre- predicting the monsoon rainfalls as an output
diction (NCEP), USA. NW India spring mean sea according to the inputs and the internal structure
level pressure anomaly was calculated from the of the network established during the learning
data of four stations, Bikaner, Jodhpur, Akola and period.
Nagpur. May Minimum temperature anomaly was The most common neural network is the feed-
calculated using the data of 7 stations (Nagpur, forward mapping network. It consists of a set of
Akola, Indore, Bhuj, Jodhpur, Rajkot and Ver- nodes and a set of interconnections between
aval). Darwin pressure tendency and Nino-3 index them. A node contains a computational element
tendency represent the ENSO phenomenon. There called neuron (Fig. 1), taking inputs from
are 3 regional parameters, ECT, NWPA and incoming interconnections (input links) and
MAYTM. providing outputs to outgoing interconnections
(output links). The units of the neural network
are arranged by layers. A unit on one layer takes
3. Brief Review of Neural Network Technique
inputs from the units on the layer below and
Neural Networks are signal processing systems feeds its output to the units on the layers above.
that attempt to emulate the behavior of biological The bottom layer is called input layer whose
nervous systems by providing a mathematical units take input from the outsided and without
model of combination of numerous basic blocks processing them distribute to the units on the
called neurons connected in a network. It is layer above. The top layer is an output layer
remotely analogous to living nervous system and whose output is the output of the neural network.
hence its name. One can think of neural networks The layers between input and output layers are
as an extended form of regression which have the called hidden layers. A pattern is de®ned as a set
properties of (i) adaptivity (ii) robustness, (iii) of input values with the related output values. A

Fig. 1. Schematic representation of a


simple neuron
258 P. Guhathakurta et al.

values for the pattern with unknown outputs (Test


period or prediction period).

4. Deterministic Neural Network Model


It has been reported (Satyan, 1988; Basu and
Andharia, 1992) that the time series of Indian
monsoon rainfall has a chaotic behavior. The
method followed by them basically consists of
the reconstruction of the attractor underlying the
time series of monsoon rainfall, estimation of
the attractor dimension and the corresponding
embedding dimension. By using standard method
they have found the embedding dimension D ˆ 7,
for proper reconstruction of the attractor. Finally
they have proposed a prediction function
Rk‡D‡1 ˆ …RK‡1 ; RK‡2 ; . . . ; RK‡D †
Fig. 2. Typical con®guration of a three layer perceptron
neural network
and approximated the prediction function by a
second degree polynomial. Higher degree approx-
imations are formidable because of the prohibi-
typical computational element takes the weighted tively large number of coef®cients.
sum of the input links and passes the result Since neural network technique is also capable
through a transfer function. to explain the nonlinearity of the time series data,
The structure of a three layer neural network is we have applied this technique to reconstruct the
shown in Fig. 2. The transfer function used here assumed deterministic dynamics of the data. A
is the sigmoidal function three layer neural network model (Guhathakurta
and Thapliyal, 1997) is used. First 50 years
1
Yˆ ; (1901±1950) of data are used for training and
1‡ e…ÿAx‡B† remaining 47 years (1951±1997) are used for
where A determines the slope of the sigmoid and forecast validation. The input layer is structured as
B is the threshold. The process of learning the continuous and discrete. For continuous input
training set of patterns means the determination layer model, continuous years of past data are
of the optimum weights which minimize the used to predict one year ahead of prediction. From
mean square error between the output in the a different approach by examining the perfor-
output layer and the desired values. Most com- mance of the continuous input layer model in the
monly used ``back-propagation learning algo- forecast validation period (1951±1997) for several
rithm'' (Rumelhart et al., 1986) is used for the number of inputs it is found that seven number
training. Initially random weights between 0.5 of inputs i.e., immediately past seven years data
are assigned to each weights as initial guesses. gives the optimum result.
The weights are learned through an iterative For the discrete input layer model, we have to
process. During learning the weights are updated introduce a gap in the data for input. Here, the
according to the formula. input consists of a set of immediately past years
of data and another set of past years of data by
New weight change ˆ Learning rate  Error leaving some years in between i.e., for predicting
‡ Momentum  Last weight change; (K‡m‡n‡l‡1) year rainfall Rk‡m‡n‡l‡1 inputs
are taken as
where learning rate and momentum rate are
required to get the convergence faster. When the Rk‡1 ; Rk‡2 ; . . . . . . Rk‡m ; Rk‡m‡n‡1 ;
network learns the training set of patterns well
enough, it can be used for determining the output Rk‡m‡n‡2 ; . . . . . . ; Rk‡m‡n‡l ;
Long Range Forecasting Indian Summer Monsoon Rainfall 259

Fig. 3. Performance of the de-


terministic neural network mod-
el during the independent period,
1951±1997

leaving n years data in between, the model is layer model during the independent period
tested by seeing the performance in the indepen- (1951±1997).
dent period (1951±97) for different combinations
RMSE ˆ 10:3% and Correlation Coefficient
of m, l (i.e., two sets of input) and n, the number
of gaps. Here m and l are varied from 2 to 9 and n ˆ 0:40:
is varied from 0 to 4. By examining several However, its performance for the 10 year period,
combinations of number of two sets and number 1988±1997 (Table 2) is quiet promising as RMSE
of skipping years, it is found that past 13 years is 8.67% and correlation coef®cient (C.C.)
(l ˆ 5, m ˆ 8, n ˆ 2) give the optimum result between the actual and predicted rainfall is 0.41.
(Guhathakurta and Thapliyal, 1997). The result is
also better than the continuous input layer model.
For both the continuous and discrete input layer
models, ®ve hidden nodes give the optimum 5. Eight Parameter Principal Component
result. Figure 3 shows the performance of the Neural Network Model
model during the independent period (1951± Since the deterministic neural network (Model I)
1997). The following results are obtained did not show useful predictive skill, we have used
regarding the performance of the discrete input eight regional and global parameters as inputs to
develop a neural network model. The eight par-
ameters used in this model are given in Table 1.
Table 2. Performance of the Discrete Input Layer Model Since some of the parameters show strong inter-
During 1988±97 correlations, we have carried out Principal
Component Analysis (PCA) of these 8 param-
Year Actual Predicted
eters. Hsieh and Tang (1998) have recom-
1988 119 98 mended the PCA method as a pre®lter for
1989 101 100 compressing the large spatial data ®eld for
1990 106 104
1991 91 95
training the neural network. PCA is a useful tool
1992 93 85 of transforming original variables into a set of
1993 100 99 uncorrelated variables. It consists of ®nding an
1994 110 107 orthogonal transformation of the original variables
1995 100 97 to a new set of uncorrelated variables, called
1996 103 96 principal components or empirical orthogonal
1997 102 112
functions which are derived in a decreasing order
260 P. Guhathakurta et al.

of importance. The components are linear combi- of data, 1958±87. In all models discussed below,
nations of the original variable but are orthogonal the models are developed using the principal
and uncorrelated to each other. Geometrically components of the period 1958±87. Using the
these linear combinations represent the selection weights/regression coef®cients thus calculated,
of a new coordinate system obtained by rotating independent forecasts are prepared for the inde-
the original system so that the new axes represent pendent period 1988±97.
the directions with maximum variability and All the eight parameters are chosen after a
provide a simple and more parsimonious descrip- careful investigation by seeing their physical
tion of the covariance structure. Other advantage linkage and degree of relationship with the
of the PCA is to reduce dimensionality of the data monsoon rainfall. The correlation coef®cients of
as the ®rst few components will account for most each parameter with monsoon rainfall for the
of the variation in the original data and by period 1958±1997 are shown in Table 1. All these
truncating the insigni®cant modes, noise level also parameters are signi®cantly correlated with the
is reduced. PCA has been carried out for 30 years monsoon rainfall. Figure 4 gives the correlation

Fig. 4. Correlation coef®cients for the 15 year


sliding window period of each of the 8
parameters with the all India monsoon rainfall
Long Range Forecasting Indian Summer Monsoon Rainfall 261

Fig. 5. a) Correlation coef®cients of each of


the 8 parameters with the monsoon rainfall of
India for the period 1958±1987 and loadings of
b) EOF1 and c) EOF2

coef®cient values for the 15 year sliding window Table 3 gives the variance explained by each of
period of each of the eight parameters with the the eigen modes. The ®rst ®ve principal compo-
monsoon rainfall plotted against the middle year nents explain 84% of the total variance and we
i.e., 1965 for the period 1958 to 1972 and so on, have used the ®rst ®ve principal components as
while Fig. 5 a gives the correlation values of each inputs in the neural network model. Figure 5b, c
of the eight parameters with rainfall for the period give the loadings of ®rst two eigen modes. The
1958±87. All the eight parameters are almost ®rst mode contributes 36% of variation. The C.C.
maintaining same degree of relationship through- and loadings of EOF1 show similar pattern among
out the period, particularly from 1980 onwards. the 8 parameters.
Therefore the choice of these parameters as A three layer neural network model with ®ve
predictors are justi®ed. inputs, three hidden nodes of the single hidden
262 P. Guhathakurta et al.

Table 3. Results of the Principal Component Analysis of the technique we have developed a multiple regres-
Predictors sion model using the same ®ve principal
Eigen Mode Eigen Value Cumulative component as predictors. Principal component
Variance data for the period 1958±1987 was used for
EOF1 2.734 34.1 developing the model. The performance of the
EOF2 1.258 49.4 Principal Component Regression Model (Model
EOF3 1.239 65.4 IV) for the independent period 1988±1997 is as
EOF4 0.824 75.7 follows.
EOF5 0.650 83.8
EOF6 0.531 90.4 RMSE ˆ 5:87% and Correlation Coefficient
EOF7 0.469 96.3
EOF8 0.291 100 ˆ 0:69:
Therefore, we have seen that improvement of
the performance of Principal Component Neural
Network Model is only slightly better compared
layer is developed to predict the monsoon rainfall to the Principal Component Regression model
of India. The most used error back propagation but is certainly appreciably better than the
algorithm (Rumelhart et al., 1986) is used for deterministic model (Model I) (RMSE ˆ 8.67%).
training. While training the network special The performances of the 8 parameter Principal
attention has been taken to avoid the local Component Neural Network Model during the
minimum (Master, 1993). The network is trained training period (1958±1987) and veri®cation
for varying number of hidden nodes. By com- period (1988±97) are shown in Fig. 6.
paring the performance of the model both in
training and testing period (1988±1997) it is
found that three number of hidden nodes gives 6. Hybrid Neural Network Model
the optimum performance. Following is the per- Chaotic behavior of the Indian summer monsoon
formance of the 8 parameter Principal Compo- rainfall time series makes it dif®cult to represent
nent Neural Network model (Model II) during it completely by a deterministic process. The
the independent period 1988±1997. Model I which is based on the deterministic
RMSE ˆ 5:74% and Correlation Coefficient process and the Model II which is based on the
ˆ 0:70: stochastic process have shown some success in
predicting summer monsoon rainfall of India. To
To compare the performance of this neural improve the performance further, we have
network model with the multiple regression combined both the models by a simple neural

Fig. 6. Performance of the 8


parameter Principal Component
Neural Network Model during
the training period (1958±1987)
and independent period (1988±
97)
Long Range Forecasting Indian Summer Monsoon Rainfall 263

Fig. 7. Performance of the hy-


brid principal component neural
network model during the train-
ing period (1958±87) and inde-
pendent period (1988±97)

Fig. 8. Schematic diagram showing the details of the models discussed

network model, called the Hybrid Neural Net- are the outputs of the earlier two models (Model
work model (Model III). In this model, we have I, based on deterministic process and Model II
considered a simple two layer neural network based on ®ve principal components), the output
without any hidden layer. The input layer of the being the rainfall. Therefore in this model there
hybrid model consists of two input nodes which are only three unknowns, namely two weights
264 P. Guhathakurta et al.

and a bias, values of which are obtained by compared with the Principal Component Regres-
training the network. The 30 years (1958±87) of sion model developed using same ®ve principal
data are used for the training. The validation is component as the inputs (Model IV in Fig. 8) and
carried out for the remaining 10 years of data the 16 Parameter Power Regression model
(1988±97). The following results are obtained (Gowariker et al., 1991) based on which India
during the forecast validation period. Meteorological Department is issuing opera-
tional long range forecasts of monsoon rainfall.
RMSE ˆ 4:93% and Correlation Coefficient
In all these cases, models were developed by
ˆ 0:80: using 30 years data (1958±1987) and forecast
The performances of the hybrid model during validations were carried out for the period 1988±
the training period 1958±87 and during the 1997. Table 4 gives forecasts based on these
veri®cation period 1988±97 are shown in Fig. 7. three models along with actual values of
It is seen that out of 10 years in the veri®cation monsoon rainfall of India and the RMSE, and
period, in 6 years (60%), forecast errors are correlation coef®cients during the independent
within one RMSE limit (4.93%) and in the period (1988±1997). The results clearly reveal
remaining 4 years, they are within 2 RMSE limit that the performance of the hybrid neural
(9.8%). Thus the hybrid model which was network model is far better than other two
developed by combining the two networks models as its RMSE is only 4.93%. The Principal
performed better than other models both during Component Regression model also performed
the training and independent periods. better than the 16 parameter power regression
The details of the development of all these model. We have also compared our results with
models discussed above are also shown schema- the model results of Goswami and Vidya (1996).
tically in Fig. 8. Their results are available for the period 1996±
1998. The independent forecasts of the neural
network model of Goswami and Vidya (1996) for
the years 1996±98 are 102%, 99.5% and 111%,
7. Comparisons of the model with other models
respectively (Goswami and Kumar, 1997). The
We have now seen that the performance of the realized rainfall of these years are 103%, 102%,
hybrid neural network model is quite encoura- and 105%, respectively. The forecasts based on
ging as the RMSE is less than 5%, whereas the the hybrid model developed by us for these years
standard deviation of all India monsoon rainfall are 103%, 96% and 105%, respectively. Thus the
is 10%. The performance of this hybrid model is present hybrid model performed comparatively

Table 4. Comparison of Different Models During the Veri®cation Period: 1988±97


Year Actual 16 Para 8 Para 8 Para Hybrid
(% of Normal) Power Principal P. C. Neural P. C. Neural
Regression Component Network Network
Model Regression Model Model
Model
1988 119 113 120 112 112
1989 101 101 102 99 99
1990 106 100 110 105 107
1991 91 101 103 101 100
1992 93 96 99 97 94
1993 100 105 100 98 98
1994 110 98 104 102 104
1995 100 103 102 101 101
1996 103 108 104 103 103
1997 102 94 92 93 96
RMSE 6.81% 5.87% 5.74% 4.93%
C.C. 0.51 0.69 0.70 0.80
Long Range Forecasting Indian Summer Monsoon Rainfall 265

better than the model of Goswami and Vidya Regression model in which the deterministic
(1996) during this small period. effects are not included forecasted much lower
(92%) rainfall in 1997. The hybrid model and the
Principal Component Regression model (Model
8. Conclusions
IV) were used for the experimental forecasts in
In this paper, we have presented the results of the 1998 by the India Meteorological Department.
models developed for Long Range Forecasting of The forecasts based on these models were 105%
Indian summer monsoon rainfall using a new and 107% respectively as against the realized
technique called Arti®cial Neural Network which rainfall of 105%. Better forecast skill from these
is currently being applied to many complex models can be achieved by using better pre-
problems because of its ability to explain the dictors. Identi®cation of better predictors and
complex behavior. We have not used either diagnostic studies of the inter-annual variability
regression analysis or nonlinear dynamics but of Indian summer monsoon are important agenda
we got the advantage of both of these in the in long range forecasting research.
neural network technique.
The discrete input layer model (Model I) is
constructed by examining all possible number Acknowledgements
and combinations of input nodes, hidden nodes, We are grateful to the Director General of Meteorology,
learning rate parameter and momentum rate to India Meteorological Department (IMD) for giving permis-
get the optimum results. The discrete input sion to publish this paper in this journal. We are also
layer model can give prediction of monsoon grateful to Dr. U. S. De, ADGM (Research) for providing
all facilities for this study and for his encouragement.
rainfall by eight months in advance. Though We also thank two anonymous referees for their con-
its performance for the period 1988±1997 is structive comments which helped to improve the quality of
good (RMSE ˆ 8.67%), its overall performance the paper.
is not so satisfactory as the model does not
include any physical mechanisms or forcings of
the inter-annual variation of Indian summer References
monsoon. Basu, S., Andharia, H. I., 1992. The chaotic time series of
The eight parameter Principal Component Indian monsoon rainfall and its prediction. Proc. Ind.
Model (Model II) performs very well for all the Acad. Sci., 101, 27±34.
10 year (1988±1997) prediction period. The Bhalme, H. N., Rahalkar, S. S., Sikder, A. B., 1987: Tropical
quasi-biennial oscillation of 10 mb wind and Indian mon-
model successfully indicated the normal and soon rainfall ± implication for forecasting. J. Climatol., 7,
excess years during the period, 1988±1997. The 345±353.
hybrid neural network model exhibits both the Bonzar, M., Lesjak, M., Mlakar, P., 1993: A neural network
characteristics of deterministic process and based method for short term predictions of ambient SO2
stochastic process by combining the two differ- concentration in highly polluted industrial areas of com-
plex terrain. Atmos. Environ., Part B: Urban Atmosphere,
ent network models. The RMSE and the correla- 27B, 221±230.
tion coef®cient between the actual and predicted Elsner, J. B., Tsonis, A. A., 1992: Nonlinear prediction,
values during the independent period indicate its chaos and noise. Bull. Amer. Meteor. Soc., 73, 49±60.
superiority over other models. We have seen that Gowariker, V., Thapliyal, V., Sarkar, R. P., Mandal, G. S.,
the performance of the hybrid model has been Sikka, D. R., 1989: Parametric and power regression
improve over the Principal Component Neural models: new approach to long range forecasting of mon-
soon rainfall in India. Mausam, 40, 115±122.
Network model and Principal Component Gowariker, V., Thapliyal, V., Kulshrestha, S. M., Mandal, G.
Regression model, because the deterministic S., Sen Roy, N. Sikka, D. R., 1991: a power regression
process is included in the model. The determi- model for long range forecast of southwest monsoon
nistic model (Model I) for recent couple of years, rainfall over India. Mausam, 42, 125±130.
especially in 1997, showed good predictive skill. Goswami, P., Srividya, 1996: A novel neural network design
for long range prediction of rainfall pattern. Curr. Sci., 70,
The year 1997 was a severe El Nino-Year. The 447±457.
Hybrid model forecast for the year 1997 was Goswami, P., Kumar, P., 1997: Experimental annual forecast
96% as against the realized rainfall of 102%. It of all India mean summer monsoon rainfall for 1997 using
is to be noted that the Principal Component a neural network model. Curr. Sci., 72, 781±782.
266 P. Guhathakurta et al.: Long Range Forecasting Indian Summer Monsoon Rainfall

Guhathakurta, P., 1998: A neural network model for short term Shukla, J., 1987: Interannual variability of monsoons. In:
prediction of surface ozone at Pune. Mausam (in press). Monsoons. (eds.) Fein, J. S., Stephens, P. L., pp. 523±548.
Guhathakurta, P., Thapliyal, V., 1997: A neural network Shukla, J., Mooley, D. A., 1987: Empirical prediction of the
model for long range prediction of monsoon rainfall over summer monsoon rainfall over India. Mon. Wea. Rev., 115,
India. Mausam (in press). 695±703.
Hsieh, W. W., Tang, B., 1998: Applying Neural Network Thapliyal, V., 1982: Stochastic Dynamic model for long
models to prediction and data analysis i Meteorology and range prediction of monsoon rainfall in Peninsular India.
Oceanography. Bull. Amer. Meteor. Soc., 79, 1855±1870. Mausam., 33, 399±404.
Krishnakumar, K., Soman, M. K., Rupakumar, K., 1995: Thapliyal, V., 1990: Long range prediction of summer
Seasonal forecasting of Indian summer monsoon rainfall: monsoon rainfall over India: Evolution and development
A review. Weather, 150, 449±467. of new models. Mausam, 91, 339±346.
Master, T., 1993: Practical Neural Network Receipes in Thapliyal, V., 1997: Preliminary and ®nal long range fore-
C‡‡. New York, Academic Press. cast for seasonal monsoon rainfall over India. J. Arid
Rajeevan, M., Pai, D. S., Thapliyal, V., 1998: Spatial and Environ., 36, 385±403.
temporal relationships between global land surface air Venkatesan, C., Raskar, S. D., Tambe, S. S., Kulkarni, B. D.,
temperature anomalies and Indian summer monsoon rain- Keshavamurty, R. N., 1997: Prediction of all India mon-
fall. Meterol Atmos. Phys., 66, 157±171. soon rainfall using error-back-propagation neural net-
Rumelhart, D., Hinton, G. E., Williams, R. J., 1986: Learn- works. Meteorol. Atmos. Phys., 62, 225±240.
ing internal representation by error propagation. In: Par- Walker G. T., 1910: Correlation in seasonal variations of
allel Distributed Processing: Exploration in the weather II. Mem. India. Met. Dep., XXI, XXII.
Microstructure of Cognition, Vol. I. Cambridge, MIT
Press, pp. 318±362. Authors' addresses: Dr. Pulak Guhathakurta, Regional
Satyan, V., 1988: Is there an attractor for the Indian summer Meteorological Centre, Civil Aerodrome, Guwahati, India,
monsoon rainfall: Proc. Ind. Acad. Sci. (Earth and Planet 780 015; Dr. M. Rajeevan and Dr. V. Thapliyal, India
Sci.), 97, 49±52. Meteorological Department, Pune-5, India, 411 005

You might also like