Characterizing and Forecasting UPLB Rainfall through Neural

Networks Approach
Brylle B. Balad-on 1, Hannah Rissah F. Abad2*, Alvin Karlo G. Tapia1 and Sharon P. Lubag1
1
Physics Division, Institute of Mathematical Sciences and Physics, University of the Philippines-Los Banos
2
Department of Mathematics, Physics and Statistics, Visayas State University
*[email protected]

Abstract
This paper presents an overview of the observed behavior of the UPLB
rainfall from 1959-2008 through neural networks approach. A 50 year daily
rainfall data, daily mean temperature, daily relative humidity and daily
sunshine duration was obtained from the Agrometeorology and Farm
Structure Division, College of Engineering and AgroIndustrial Technology.
This study introduces the effect of forecasting UPLB rainfall by the
conventions in most of the Neural Network studies based on three different
time horizons; 1) the 366 days basis, 2) days in a month basis, and 3)
monthly basis. A nonlinear function was approximated showing rainy season
from the month of June to November and dry season throughout the rest of
the year. The days in a month basis showed that Neural Network performed
better in forecasting dry season than rainy season. The monthly basis forecast
performed best with almost the same peaks as the actual values.
Keywords: Neural Networks, Rainfall Forecasting, Rainfall Characterizing
1. Introduction
Understanding the patterns and trends, cause and effect relationships and characteristics of rainfall are very
important for some institutions. Having accurate information about rainfall will aid in the planning and
management of water resources, reservoir operation and flood prevention [1]. Moreover quantitatively
predicting rainfall will aid in crop planting selections [2]. However rainfall prediction is one of the most
challenging task since rainfall are highly non-linear and complicated occurrences, which need advance
computer-based simulation and modelling for accurate forecast [1].
So far according to Zwiers and Von Storch (2004), long-term climate prediction using numerical models has
not revealed useful performance [3]. Hsieh and Tang [4] define the Neural Network (NN) method as a
computational technique that “nonlinearly relates a set of variables to another set of variables. Also, they added
that neural network technique has been successful in every time series prediction and analysis competition.
When the time series is noisy, the underlying dynamical system is nonlinear and temporal dependencies span
long time intervals, Artificial Neural Networks models frequently outperform standard linear techniques [5].
Lee, Cho and Wong (1998) observed that the artificial neural networks produced good forecasts while the linear
models produced poor forecasts [2]. NN ability to cope with nonlinear data, speed of computation, the learning
capability and the accuracy made them valuable tools of forecast [6]. NN also is suitable in modeling dynamic
systems on a real-time scenario and has the capability to make generalization from unseen data [7]. NN
modelling generally aims to develop functional tools that can yield more accurate predictions and more
dependable timings and forewarnings of dangerous flood events [8]
Several studies have utilized the NN approach in characterizing and forecasting rainfall data such as the
work by Mar and Naing in Myanmar [9], Iseri, et. al. [3] in Fukuoka City, Japan, Abraham [10] and his
associates in India, Chantasut, Charoenjit, and Tanprasertin [2] in Chao Phraya River, Junior and Barreto [5] in
Fortaleza, Brazil. This research addresses the need of a general pattern of UPLB rainfall in 50 years in the field
of agriculture using a more advanced computational tool which is the neural networks. Hence, the objective of
this study is to characterize and forecast UPLB rainfall through the neural network approach.

2. Methodology
A set of data was gathered from Agriculture and Farm Structure Division (AFSD), College of Engineering
and AgroIndustrial Technology, University of the Philippines Los Baños (UPLB). The actual observation of the
data was done by the National Agromet Station in UPLB with a geographical coordinate location of 14° 11' N
and elevation of 21.7 meters. The data is a daily observation since January 1, 1959 to August 31, 2009. Among
these data are the UPLB daily rainfall, UPLB mean temperature, UPLB sunshine duration and UPLB relative

32nd Physics Congress of the Samahang Pisika ng Pilipinas

University of the Philippines Diliman
17 – 20 October 2014
© SPP
32ⁿᵈ Physics Congress of the Samahang Pisika ng Pilipinas
1
National Institute of Physics, University of the Philippines Diliman, Quezon City, Philippines
17-20 October 2014
SPP2014-PB-01-1
humidity. The unit of measurement for the UPLB rainfall data is millimeters with a precision of one tenth of a
millimeter. A flowchart of the methodology is shown in Figure 1.
In figure 2, a NN architecture is basically composed of an input-
connections-output relationships. This connections are called the
weights. The weights of a neural network is adjusted for it to learn the
input-target relationship and produce its own input-output trend by
the result of comparison of the output and target.

Figure 2: A diagram of how a neural network typically train its input

relate it with a set of target to produce a specific output [12]

The weights and biases were repeatedly adjusted in order to

minimize the error of the output values to the target values in training
of the NN. The best combination of variables to forecast the UPLB
rainfall was determined using the least error value using the  
performance function 'mse' or mean square error. The input variable
predictors of rainfall were found out to be the daily mean
temperature, daily relative humidity, daily sunshine duration and
previous day rainfall amount. Feed forward backpropagation neural
network architecture 'newff' was chosen which was widely used by
researchers in many applications in different fields. The feed forward
Figure 1: Flowchart of the methodology backpropagation minimizes the error function using the first
derivative or second derivative optimization [11]. The rainfall data
was used as a time series using the feed forward backpropagation neural network with a 50 years (1959-2008)
data record to forecast the actual rainfall.

3. Results and Discussions

A nonlinear function was approximated using NN to characterize the 50 years daily UPLB rainfall data as
shown in figure 3. The nonlinear function approximation has shown that rainy season occurs in the months of
June to November and dry season follows throughout the rest of the year in UPLB. Moreover from the set of
data obtained, the average mean temperature increased in UPLB since 1982 and this has brought a major impact
on the increased average amount of rainfall.

Figure 3: The nonlinear function approximation of UPLB rainfall from 1959-2008

32nd Physics Congress of the Samahang Pisika ng Pilipinas

University of the Philippines Diliman
17 – 20 October 2014
© SPP
32ⁿᵈ Physics Congress of the Samahang Pisika ng Pilipinas
2
National Institute of Physics, University of the Philippines Diliman, Quezon City, Philippines
17-20 October 2014
SPP2014-PB-01-2
 The forecast values of the NN architecture was evaluated by comparing it from the past records. Moreover,
forecasting the UPLB rainfall was done based on three time horizons; the 366 days (figure 4), days in month
(figure 5-6) and monthly basis (figure 7) as presented below.

Figure 4: The 2008 actual amount of rainfall versus the forecasted amount of rainfall

Figure 5: Forecasted values by neural network and actual values of UPLB rainfall in February 2008

Figure 6: Forecasted values by neural networks and its actual values of UPLB rainfall in July 2008

Figure 7: Forecasted monthly values of rainfall in 2008 and its actual values

32nd Physics Congress of the Samahang Pisika ng Pilipinas

University of the Philippines Diliman
17 – 20 October 2014
© SPP
32ⁿᵈ Physics Congress of the Samahang Pisika ng Pilipinas
3
National Institute of Physics, University of the Philippines Diliman, Quezon City, Philippines
17-20 October 2014
SPP2014-PB-01-3
 The 366 days forecast (Figure 4) shows the inability of neural network to consider extreme conditions of
UPLB rainfall. Meanwhile, the days in a month basis shows that neural network better perform in forecasting
dry season than rainy season (Figure 5-6). The monthly basis forecast performed best with the same peaks as
the actual values. In this part extreme condition, outliers and noisy data were in a minimum (Figure 7). Taking
into consideration the results of Mar and Naing in Myanmar [9], Iseri, et. al. [3] in Fukuoka City, Japan,
Abraham [10] and his associates in India, Chantasut, Charoenjit, and Tanprasertin [2] in Chao Phraya River,
Junior and Barreto [5] in Fortaleza, Brazil a neural network model was established only for the monthly
rainfall prediction.

4. Conclusion
A nonlinear function approximation and forecasting of UPLB daily rainfall was done through neural
networks. Based on the results of function approximation, UPLB experiences dry season in months of December
to May and rainy season from June to November.
Furthermore, the UPLB rainfall in 2008 was forecasted using the Neural Networks approach in three parts;
366 days, days in a month and monthly basis. The 366 days forecast shows the inability of neural network to
consider extreme conditions of UPLB rainfall. Meanwhile, the days in a month basis shows that neural network
performed better in forecasting dry season than rainy season. Forecasting the amount of rainfall in UPLB is
more applicable using the monthly basis.

References
1. Nayak, Mahapatra, and Mishra. A Survey on Rainfall Prediction using Artificial Neural Network.
International Journal of Computer Applications (0975 – 8887).Volume 72– No.16, June 2013
2. Chantasut, Charoenjit, and Tanprasert. Predictive Mining of Rainfall Predictions Using Artificial Neural
Networks for Chao Phraya River. Proceedings of Joint Conference The 4th International Conference of The
Asian Federation ofInformation Technology in Agriculture and The 2nd World Congress on Computers in
Agriculture and Natural Resources, August 9-12, 2004, Bangkok, Thailand, pp. 117-122.
3. Iseri, et.al. Medium Term Forecasting of Rainfall using Artificial Neural Netwo`rks. International Congress
on Modelling and Simulation: Modelling and Simulation Society of Australia and New Zealand, / Andre Zerger
and Robert M. Argent (eds.): pp.1834-1840. December 2005.
4. Hsieh, W.W. and Tang, B. Applying Neural Network Models to Prediction and Data Analysis in Meteorology
and Oceanograph. Bulletin of the American Meteorological Society 9(79):1855-1869pp. 1998.
5. Junior and Barreto Multistep-Ahead Prediction of Rainfall Precipitation Using the NARX Network.
researchgate.net.
6. Sorin Vlad, On the Prediction Methods Using Neural Networks. citeseerx.ist.psu.edu
7. Bustami, et.al. Artificial Neural Network for Precipitation and Water Level Predictions of Bedup River
IAENG International Journal of Computer Science, 34:2, IJCS 34-2-10. Nov 17, 2007.
8. Abrahart, Heppenstall, and See. Timing error correction procedure applied to neural network rainfall–runoff
modelling. Hydrological Sciences–Journal–des Sciences Hydrologiques, 52(3), 415-435. June 2007.
9. Mar, K. W. and Naing, T. T. Optimum Neural Network Architecture for Precipitation Prediction of Myanmar.
World Academy of Science, Engineering and Technology. 2008.
10. Abraham, Steinberg and Philip, Rainfall Forecasting Using Soft Computing Models and Multivariate
Adaptive Regression Splines. meghnad.iucaa.ernet.
11. Elshafie, A. et al. Nonlinear prediction model for scour and air entrainment based static neural network
approach”. European Journal of Scientific Reasearch. EuroJournal Publishing Inc. 400-416pp. 2009.”
12. Demuth, et al. 2009. Neural Network ToolboxTM 6 User's Guide.

32nd Physics Congress of the Samahang Pisika ng Pilipinas

University of the Philippines Diliman
17 – 20 October 2014
© SPP
32ⁿᵈ Physics Congress of the Samahang Pisika ng Pilipinas
4
National Institute of Physics, University of the Philippines Diliman, Quezon City, Philippines
17-20 October 2014
SPP2014-PB-01-4