ISSN: 2454-9924

NEURAL NETWORK BASED RAINFALL

PREDICTION SYSTEM
Dr.B.Sowmya1, Mrs.A.Devasena2, Mrs.S.Esther Gandhimathi3, Ms.P.Dharani Devi4
1 2,3
Professor, Associate Professor, 4JRF, Department of ECE, Dhanalakshmi College of Engineering, Tambaram,
Chennai - 601 301. Tamil Nadu

ABSTRACT: Agricultural production in Kancheepuram ambiguous information (Nourani et al., 2008).Neuro-fuzzy has
district depends primarily on rainfall. The amount, been used in rainfall-runoff modeling by (Nasr and Bruen,
frequency and intensity of rainfall vary from place to place 2008; Remesan et al.,2009), hydrological time series
and time to time. Precise knowledge of rainfall is essential prediction (Nayak, 2004;Firat, 2007, 2008) prediction of water
for planning its full utilization. With the rainfall data sets level in reservoir (Chang, 2006); reconstructing missing flow
recorded all year round, rainfall prediction with predictive data, suspended sediment (Kisi, 2005, 2009), evaporation
tools is possible. An approach based on neural network for (Moghadamnia et al., 2009), river flow estimation (Firat,
rainfall forecasting in Kancheepuram district is carried out. 2007; El-Shafie, 2006).
The prediction of rainfall has been attempted using feed In recent years, ANNs have been used intensively for
forward neural networks. The regional rainfall over prediction and forecasting in a number of water-related areas,
Kancheepuram region is studied first and then used to including water resource study (Najah et al., 2009; Ahmed et
forecast rainfall. The patterns of different days, months and al., 2009; El-Shafie et al., 2007; El-Shafie et al., 2008; El-
years are combined, which are then filtered to clean noise. Shafie et al., 2009(a), El-Shafie et al., 2009(b); El-Shafie et
The trained system will predict rainfall on daily basis with al., 2010), prediction of evaporation (Sudheer et al., 2002),
acceptable accuracy. The performance of the Prediction hydrograph simulator (Deka and Chandramouli, 2005; Lange,
system is evaluated by the quality measures. 1999), rainfall estimating (Hsu et al., 1999; Lin and Chen
2005; Luk et al., 2001).Among ANN techniques, Multi-
1. INTRODUCTION Layered Perceptron Network (MLPN), Elman Recurrent
Neural Networks (ERNN) are commonly used. Though many
rainfall prediction systems have been developed, they do not
Agricultural production in Kancheepuram district depends provide the wide dynamic range. They are not adaptive in case
primarily on rainfall. The amount, frequency and intensity of non linear data.
vary from place to place and time to time. Precise knowledge
of rainfall is essential for planning its full utilization. With the
rainfall data sets recorded all year round, rainfall prediction 3. PROPOSED SYSTEM
with predictive tools is possible. An approach based on neuro
system for rainfall forecasting in Kancheepuram district is The rainfall data of Kathupakkam, Kancheepuram station
carried out. The prediction of rainfall has been attempted has been collected from Meteorological and Oceanographic
using neural networks. The regional rainfall over Satellite Data Archival Centre (MOSDAC).The regional
Kancheepuram region has been studied first and then used to rainfall over Kancheepuram region has been studied first and
forecast rainfall. The patterns of different days, months and then used to forecast rainfall. The hourly data acquired has
years are combined, which are then filtered to clean noise. been converted to daily data. From the daily data monthly
Fuzzy clustering is used for data selection for training the rainfall data has been obtained. The seasonal rainfall data is
neural and Evolutionary Neuro Fuzzy Prediction System determined from the monthly data.
(ENFPS). After the cleaning process the patterns are chosen The patterns of different days, months and years are
for training the Prediction System. The trained system will combined, which are then filtered to clean noise. Fuzzy
predict rainfall on daily, monthly, seasonal basis with clustering is used for data selection for training the neural and
acceptable accuracy. Post processing of predicted rainfall is Neuro Prediction System. After the cleaning process the
carried out to reduce bias and to quantify uncertainty. patterns are chosen. The Prediction System has been trained
with the selected patterns. The trained system will predict
2. EXISTING SYSTEM rainfall on daily, monthly, seasonal basis with acceptable
accuracy. Post processing of predicted rainfall is carried out
to reduce bias and to quantify uncertainty due to Southern
In hydrological forecasting, the artificial neural network is Oscillation Index, El-Nino etc. The performance of the
applied for rainfall-runoff modeling (Remesan and Shamim, prediction system is evaluated by the quality measures.
2009; Junsawang et al., 2007), light penetration estimation in Separate Graphical User Interface (GUI) is created for daily,
reservoirs (Soyupak et al., 2007), stream flow prediction monthly and seasonal prediction systems. Both Automatic and
(Chang, 2006; Chen et al., 2006; Kisi, 2004a; Cigizoglu and manual entry of rainfall predictors is available in the GUI. A
Kisi, 2005); Fuzzy based model could be used to model the separate GUI is created for evaluating the performance of
process behaviors even with incomplete and imprecise or prediction system. The performance of the prediction systems

1
 ISSN: 2454-9924

is evaluated by the quality measures, viz. error range and

regression value. The inputs required for the Simple daily rainfall
prediction are Average temperature of the previous day in oC,
4. METHOD OF ANALYSIS Average Wind Speed of the previous day in knots, Wind
The flow diagram of the rainfall prediction system Direction in degrees, Average Humidity in %, Atmospheric
developed is shown in Figure 4.1. Any forecasting system Pressure of the previous day in Pascal’s and Average Sunshine
requires predictors and a predictand. Predictand is a variable Duration of the previous day in Hours are given as input to the
which has to be predicted. Predictor is a variable value which prediction system.
can be used to forecast the value of another variable. In the The output of the NNBPS system is the rainfall of the day
rainfall prediction system the predictand is the amount of in mm. Initially, a set of 1657 input and output data are used to
rainfall. The predictors are identified by correlating the train the system. If a new set of input is given to the trained
parameters with the amount of rainfall. system then it will predict the rainfall of the day accurately.

NNFPS for Daily Rainfall Prediction with Refined Mean

Data Data
The inputs required for the Mean daily rainfall prediction
are Previous Day Rainfall Average rain fall in mm, Average,
Data Preprocessing Minimum and Maximum temperature of the previous day in
o
C, Average Wind Speed of the previous day in knots, Wind
Dataselection by Fuzzy Clustering Direction in degrees, Average Humidity in % and Average
Sunshine Duration of the previous day in Hours are given as
input to the prediction system. Atmospheric Pressure of the
Neural Network Based Prediction previous day in Pascal’s is directly given as input instead of
System Training and Testing Instead of Change in Atmospheric Pressure,

Postprocessing
Minimum Temperature

Rainfall Prediction System

Maximum Temperature
Rainfall forecast
Average Temperature

Figure 4.1 Neural Network Based Prediction System (NNBPS) Average Wind Speed
Daily Rainfall
Wind Direction
Rainfall Predictors
Change in Atmospheric Pressure
The factors used for determining rainfall are,
 Wind Speed Humidity

 Change in Atmospheric Pressure Sunshine

 Temperature Previous Day Rainfall
 Wind Direction
 Cloud over Figure 4.3 ENFPS for Daily Rainfall Prediction
 Relative Humidity
 Sunshine Duration The output of the NNBPS system is the rainfall of the day in
 Indian Ocean Dipole mm. Initially, a set of 1657 input and output data are used to
 Southern Oscillation Index (SOI) train the system. If a new set of input is given to the trained
 El Niño Modoki Index system then it will predict the rainfall of the day accurately.

NNFPS for Daily Rainfall Prediction with Refined NNFPS for Daily Rainfall Prediction with Proposed Data
Original Data
Rainfall Prediction System

Temperature Minimum Temperature

Rainfall Prediction System

Maximum Temperature
Average Wind Speed
Average Temperature

Wind Direction Daily Rainfall

Average Wind Speed
Daily Rainfall
Wind Direction
Atmospheric Pressure
Change in Atmospheric Pressure

Humidity Humidity

Sunshine
Sunshine
Previous Day Rainfall

Figure 4.2 ENFPS for Daily Rainfall Prediction with Refined

Data
Figure 4.4 NNBPS for Daily Rainfall Prediction
2
 ISSN: 2454-9924

The inputs required for the Proposed daily rainfall

prediction are Previous Day Rainfall Average rain fall in mm, The hidden layer neurons of a Feed Forward Neural Network
Average, Minimum and Maximum temperature of the take sigmoid activation function. The sigmoid activation
previous day in oC, Average Wind Speed of the previous day function is given by,
in knots, Wind Direction in degrees, Change in Atmospheric 1
Pressure of the previous day in Pascal’s, Average Humidity in F ( n) 
1  exp(  * n)
% and Average Sunshine Duration of the previous day in
Hours are given as input to the prediction system. Linear activation function for the output layer has been
The output of the NNBPS system is the rainfall of the day used. These neurons give the output directly proportional to
in mm. Initially, a set of 1657 input and output data are used to their input. i.e. the input and output have linear relationship.
train the system. If a new set of input is given to the trained Back propagation learning algorithm is used for training.
system then it will predict the rainfall of the day accurately. Figure 4.9 shows the sigmoid function. The linear activation
function is shown in Figure 4.10.
Neural Network as Supervised Classification System The network is assumed to have n inputs denoted by xi
Neural networks can also be used as supervised where ivary from 1 to n, m outputs denoted by yk where k
classification system. The operation of a neural network as varies from 1 to m. There are n input neurons and m output
supervised classification system comprises of the following neurons. The hidden layer has p neurons.
steps. The connection between layers and their weights are
shown in figure 4.10. The weights of the input to hidden layer
Training of the system connections are marked as v and the weights of hidden layer to
output layer links are marked as w.

Figure 4.10 Purely Linear Function

Figure 4.8 Training of Neural Network
5. RESULTS
Testing of the system
The prediction of rainfall has been attempted using
Training is the process of adjusting the weights of links in
different techniques including artificial neural networks and
such a way that a particular input leads to specific target
Neuro Fuzzy systems separately. The regional rainfall over
output. Figure 4.8 shows the training of a neural network.
Kancheepuram region has been studied first and then used to
forecast rainfall. The data are preprocessed to suit the
Feed Forward Neural Network
seasonal, monthly and daily rainfall predication systems. As
A Feed forward neural network is a biologically inspired
ENFPS failed to predict the rainfall correctly, the prediction
classification network. In this, information is passed in the
by three neural networks is compared on the basis of quality
forward direction only (Simon Haykins 1999). In a
measures.
Feedforward network neurons are arranged in layers, with the
Quality Measures for Performance Analysis
first layer taking in inputs and the last layer producing outputs.
The quality measures used are Nash-Sutcliffe Coefficient,
The middle layers which have no connection with the external
Gamma Coefficient, Correlation Coefficient, Spearman
world are called hidden layers. Each neuron in one layer is
Correlation coefficient and Root Mean Square Error.
connected to every neuron in the next layer. These networks
Nash-Sutcliffe Coefficient
are known as feed forward networks, since the information is
The Nash-Sutcliffe Coefficient (NSC) is a normalized
constantly "fed forward" from previous layer to the next layer.
statistic that determines the relative magnitude of the residual
There is no connection between neurons in the same layer.
variance between actual value and the data value predicted by
the model. It is given by,
(𝐴𝑐𝑡𝑢𝑎𝑙𝑉𝑎𝑙𝑢𝑒 − 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑𝑉𝑎𝑙𝑢𝑒)2
𝑁𝑆𝐸 = 1 −
(𝐴𝑐𝑡𝑢𝑎𝑙𝑉𝑎𝑙𝑢𝑒 − 𝑚𝑎𝑒𝑎𝑛 𝐴𝑐𝑡𝑢𝑎𝑙𝑉𝑎𝑙𝑢𝑒 )2
 Nash-Sutcliffe efficiencies range from - ∞ to 1.
Essentially, the closer to 1, the more accurate the
model is.
 NSE = 1, corresponds to a perfect match of predicted
data and the actual data.
 NSE = 0, indicates that the model predictions are as
Figure 4.9 Sigmoid Function accurate as the mean of the actual data.
3
 ISSN: 2454-9924

 -∞< NSE < 0, indicates that the observed mean is There are five different prediction systems with three different
better predictor than the model. neural networks. Their training patterns are explained here.
Figure 5.1 shows the artificial Feed forward neural
Gamma Coefficient network used for predicting daily rainfall with refined original
Gamma Coefficient is a measure of relationship between data. It has three layers. They are input layer, output layer and
variables. Gamma value can vary from 0.0 to ±1.0. It hidden layer. There are 6 input neurons for the six prediction
indicates the strength as well as the direction of the variables. The hidden layer has 40 neurons and the activation
relationship between the variables. The gamma coefficient is function is sigmoid function. There is one neuron in the output
given by layer.
(𝑁𝑆 − 𝑁𝐷 ) Four different plots can be opened from the training
𝐺𝑎𝑚𝑚𝑎 = window. They are
(𝑁𝑆 + 𝑁𝐷 )
NS is the same ordered pair in the variables  Performance plot
ND is the inverse ordered pair in the variables  Training state plot
NS shows positive relationship between variables. i.e. if  Error histogram plot
variable A is high then B is also high and vice versa. N D shows  Regression plot
negative relationship between variables. i.e. if variable A is The performance plot is the graph between the values of
high then B is low and vice versa. the performance function and the iteration number.
Performance Curve for Feed forward Neural Network with
Correlation coefficient daily refined data is shown in Figure 5.2. The performance of
Correlation coefficient is a measure of strength and training, validation, and test states are plotted. The iteration at
direction of relationship between variables. Its value varies which the validation performance reached a minimum is
between -1 to +1. Positive correlation between variables result indicated as the best epoch. It occurs at epoch 0. After
in positive values and negative correlation between variables reaching the best epoch the training is continued for 6 more
result in negative values. +1 shows a perfect positive fit. The iterations before the stopping.
Correlation coefficient is given by,
Best Validation Performance is 111035.3996 at epoch 0
Correlation Coefficient 10
6

n xy − ( x)( y) Train
= Validation
Test
2 2
n( x 2 ) − ( x) n( y 2 ) − ( y) Best
Mean Squared Error (mse)

5
where n is the number of data pairs. 10

Spearman correlation coefficient

The Spearman correlation coefficient is also known as the 4
10
Pearson correlation coefficient. It is a measure of correlation
between the ranked variables. If X and Y are the variables,
they are converted to ranks x and y. Then, the spearman
coefficient is calculated from, 10
3

6 𝑑2 0 1 2 3 4 5 6
𝑆𝑝𝑒𝑎𝑟𝑚𝑎𝑛𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡 = 1 − 6 Epochs
𝑛(𝑛2 − 1)
where n is the number of data pairs. Its value varies between - Figure 5.2 Performance Curve for Feed forward Neural Network
(daily refined data)
1 to +1. Positive correlation between variables result in
positive values and negative correlation between variables
result in negative values. +1 shows a perfect positive fit. 10
6
Gradient = 168081.0194, at epoch 6
gradient

5
10
Root Mean Square Error
4
10
𝑛 2
𝑖=1 𝑥𝑖 − 𝑦𝑖 Mu = 1, at epoch 6
𝑅𝑀𝑆𝐸 = 10
5

𝑛 0
mu

10

-5
10

Validation Checks = 6, at epoch 6

10
val fail

0
0 1 2 3 4 5 6
6 Epochs

Figure 5.1 Trained Feed forward Neural Network (daily Figure 5.3 Training State for Feed forward Neural Network
refined data) (daily refined data)

4
 ISSN: 2454-9924

Error Histogram with 20 Bins output for the training inputs. The second plot shows the
1400 network output for the inputs during validation. The third plot
Training
1200 Validation shows the network output for the inputs during testing
Test
Zero Error
1000 Performance Evaluation of Rainfall Prediction Systems
Instances

800
TABLE 5.1
600
FFNN
400
Regression
200 Dataset Error Range
Value
0 Refined
-1324
-1171
-1018

1123
1276
1429
1582
-865

52.53
205.5
358.4
511.3
664.2
817.2
970.1
-712.1
-559.2
-406.2
-253.3
-100.4

-1300 to 1600 0.74845

Original
Errors = Targets - Outputs Refined
Figure 5.4 Error Histogram for Feed forward Neural Network -1 to 1 0.9494
Mean
(daily refined data)
Refined
-2 to 1 0.9542
The training state plot is shown in Figure 5.3. The training Proposed
state plot shows the progress of other training variables, such
as the gradient magnitude, the number of validation checks, The fourth plot shows the trained network output to all of
etc. Gradient is the back propagation gradient shown in the inputs in the data set. The regression value R is between
logarithmic scale. 0.73 and 0.76. It shows poor matching between the network
The number of validation checks denotes the number of outputs and the targets.
consecutive iterations that the performance of validation failed Table 5.1 shows the error range and regression values
to decrease. If this number reaches default value, the training obtained for rainfall prediction with different neural network
stops. with different sets of data. From the table it can be seen that
The distribution of the network errors is shown in the refined proposed data gives best regression value and the error
error histogram plot. It is shown in Figure 5.4. The blue bars range is lowest for refined mean data.
Training data are represented by blue bars. Validation data are
represented by the green bars and the testing data are 6. CONCLUSIONS
represented by the red bars. It can be seen that the maximum
errors fall between -1300 and 1600. The rainfall predication system using feed forward neural
network is trained with three types of refined prediction data
sets. In the first case, original values of prediction parameters
are directly given. The next system is trained with refined
mean data. Lastly the system is trained with refined proposed
data. The results show that the best predictions are obtained
with refined proposed data.

ACKNOWLEDGEMENT

This research work is sponsored by NARL, Gadanki

under the ISRO - RSPOND scheme.

REFERENCES

1. Ahmed AN, Elshafie A, Karim O, Jaffar O (2009).

Evaluation the Efficiency Of Radial Basis Function
Neural Network For Prediction Of Water Quality
Parameters. Eng., Intell. Syst., 17(4): 221-231.
2. Bezdek J. C., 1981, “Pattern Recognition with Fuzzy
Figure 5.5 Regression for Feed forward Neural Network Objective Function Algorithms”, Plenum Press, New
(daily refined data) York.
3. Chen SH, Lin YH (2006). The strategy of building a
Regression plot is shown in Figure 5.5 Regression plot flood forecast model by neuro-fuzzy network.
shows the relationship between the network outputs and the Hydro.Proc., 20(7): 1525-1540.
targets. If the training is perfect, the network outputs and the 4. Cigizoglu HK, Kisi O (2005). Flow prediction by three
targets are exactly equal. The first plot shows the network back propagation tech., using k-fold partitioning of
5
 ISSN: 2454-9924

neural network training data. Nordic Hydrol., 36(1): 49- application to rainfall estimation. Water Resour.
64 Res.,35(5):1605-1618.
5. Deka P, Chandramouli V (2005). Fuzzy neural network 18. http://www.wikipedia.org
model for hydrologic flow routing. J. Hydrol. Eng.,
19. http://www.jamstec.go.jp
10(4): 302-314.
6. Dunn.J.C, 1973, “A Fuzzy Relative of the ISODATA 20. Jang, SR (1993). ANFIS: Adaptive-network-based
Process and Its Use in Detecting Compact Well- fuzzy inference system. IEEE Trans. Syst. Man
Separated Clusters”, Journal of Cybernetics Vol.3, Cybern., 23(3): 665-685.
pp.32-57. 21. Jatinder N.D Gupta, Randall S Sexton S
7. Duraisamy, N. Devarajan, D. Somasundareswari, A. (1999).Comparing back propagation with a genetic
Antony Maria Vasanth, S.N. Sivanandam, 2007,“Neuro algorithm for neural network training, Omega, 27(6):
fuzzy schemes for fault detection in power 679-684
transformer”, Applied Soft Computing Vol.7, 2007, pp. 22. Kisi O (2003). Daily River Flow Forecasting using
534–539. Artificial Neural Networks and Auto Regression
8. EdvinAldrian, YudhaSetiawanDjamil, (2008), Model. Turkish J. Eng. Environ. Sci.,29: 9-20.
“Application of multivariate ANFIS for daily rainfall 23. Kisi O (2004a). River flow modeling using artificial
prediction : influences of training data size”, neural networks. J.Hydrol. Eng. ASCE, 9(1): 60-63.
Makara,SainsVol 12(1). 24. Kisi O (2005). Suspended sediment estimation using
9. April 2008.El-Shafie A , Najah AA , Karim O (2009)b. neuro-fuzzy and neural network approaches/Estimation
Application of neural network for scour and air des matières en suspension par des
entrainment prediction. Inter. Conference onComputer approchesneurofloues et à base de réseau de neurones.
Techno. Dev., Art. no. 5360151: 273-277 Hydro. Sci. J., 50(4): 683-696.
10. El-Shafie A, Alaa, Noureldin A, Mohd RT (2009a). 25. Kisi O, Haktanir T (2009). Adaptive neuro-fuzzy
Enhancing Inflow Forecasting Model at Aswan High computing tech., for suspended sediment estimation.
Dam Utilizing Radial Basis NeuralNetwork and Adv. Eng. Softw., 40(6): 438-444.
Upstream Monitoring Stations Measurements. Water 26. Lange NT (1999). New mathematical approaches in
Resour. Manage., 23(11): 2289-2315. hydro. Modeling an application of artificial neural
11. El-Shafie A, Noureldin A (2010). Generalized versus networks. Phys. Chem. Earth, (B); 24(1–2):31–5.
non-generalized neural network model for multi-lead 27. Lin GF, Chen LH (2005). Application of an artificial
inflow forecasting at Aswan High Dam. Hydro., Earth neural network to typhoon rainfall forecasting. Hydro.
Syst. Sci. Discuss., 7(5): 7957-7993. Proc., 19:1825-37.
12. El-Shafie A, Noureldin AE, Taha MR, Basri H (2008). 28. Luk KC, Ball JE, Sharma A (2001). An application of
Neural network model for Nile River inflow forecasting artificial Neural networks for rainfall forecasting. Math.
based on correlation analysis of historical inflow data. Comput. Model., 33: 683-693.
J. Appl. Sci., 8(24): 4487-4499 29. Nasr A, Bruen M (2008). Development of neuro-fuzzy
13. El-Shafie A, Reda TM, Noureldin A (2007) A Neuro- models to account for temporal and spatial variations in
Fuzzy Model for Inflow Forecasting of the Nile River a lumped rainfall-runoff model. J. Hydro., 349(3-4):
at Aswan High Dam. Water Resources Manage., 21(3): 277-290.
533-556. 30. Nayak P C, Sudheer K P (2004). A neuro-fuzzy
14. Firat M, Güngör M (2007). River flow estimation using computing technique for modeling hydro. time series. J.
adaptive neuro fuzzy inference syst. Math. Comput. Hydro., 291(1-2): 52-66.
Simul., 75(3-4): 87-96. 31. Nourani V, Mogaddam AA, Nadiri AO (2008a). An
15. Firat M, Güngör M (2008). Hydro., time-series ANN-based model for spatiotempral groundwater level
modelling using an adaptive neuro-fuzzy inference syst. forecasting. Hydro. Proc., 22: 5054-5066.
Hydrol. Proc., 22(13): 2122-2132. 32. Remesan R, Shamim MA (2009). Runoff prediction
16. Hong-Rui Wang Li Yang Li-Xin We, 2007, “Fuzzy- using an integrated hybrid modelling scheme. J. Hydro.,
neuro position/force control for robotic manipulators 372(1-4): 48-60.
with uncertainties”, Soft Computing, 2007 vol. 11, pp. 33. Soyupak S, Karaer F (2007). Adaptive neuro-fuzzy
311–315. inference technique for estimation of light penetration
17. Hsu KL, Gupta HV, Gao X, Sorooshian S (1999). in reservoirs. Limnology, 8(2): 103-112.
Estimation of physical variables from multichannel 34. Sudheer KP, Gosain AK, Rangan DM, Saheb SM
remotely sensed imagery using a neural network: (2002). Modelling evaporation using an artificial neural
network algorithm, Hydro. Process. lfi, pp. 3189-3202.