Improvement of Daily Temperature Prediction Model for Northern Thailand using Artificial Neural Networks

Abstract

Developed in Northern Thailand, Artificial Neural Network (ANN) model for daily temperature prediction was to evaluate the performance of the developed model with the data from 24 northern meteorological stations from 1980-2011 comparing with 365 days of 2014 weather data. The model performance equals accuracy of the predicted temperature on previously chosen days. The developed models consist of daily maximum temperature forecast models (Tmax models) and daily minimum temperature forecast model (Tmin models). These models forecast ahead 24-72 hours. Performance analysis shows MAE correlation as following. For 24-hours forecast, the MAE was in the range of 0.91-1.20 for Tmin models and 1.01-1.82 for Tmax models. For 48-hours forecast, the MAE was ranging 1.06-1.47 for Tmin models and 1.22-1.81 for Tmax models while for 72-hours forecast, it varied 1.14-1.81 for Tmin models and 1.29-1.99 for Tmax models. Moreover, the R-squared showed good relationship between predicted temperature and actual temperature. Therefore, the correction of 24 hours forecast model is better than 48 hours and 72 hours, respectively.

Full text

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
Improvement of Daily Temperature Prediction Model for
Northern Thailand using Artificial Neural Networks
Sukrit Kirtsaeng*, Pattara Sukthawee, Banluesak Khosuk,
Fatah Masthawee and Nuttapong Pantong
Meteorological Development Bureau, Thai Meteorological Department, Bangkok 10260, Thailand
*e-mail: sukritk@hotmail.com
Abstract
Developed in Northern Thailand, Artificial Neural Network (ANN) model for daily
temperature prediction was to evaluate the performance of the developed model with the data
from 24 northern meteorological stations from 1980-2011 comparing with 365 days of 2014
weather data. The model performance equals accuracy of the predicted temperature on
previously chosen days. The developed models consist of daily maximum temperature
forecast models (T
max
models) and daily minimum temperature forecast model (T
min
models).
These models forecast ahead 24-72 hours. Performance analysis shows MAE correlation as
following. For 24-hours forecast, the MAE was in the range of 0.91-1.20 for T
min
models and
1.01-1.82 for T
max
models. For 48-hours forecast, the MAE was ranging 1.06-1.47 for T
min
models and 1.22-1.81 for T
max
models while for 72-hours forecast, it varied 1.14-1.81 for T
min
models and 1.29-1.99 for T
max
models. Moreover, the R-squared showed good relationship
between predicted temperature and actual temperature. Therefore, the correction of 24 hours
forecast model is better than 48 hours and 72 hours, respectively
.
Introduction
In this research, ANN (Artificial Neural Network) model is applied for developing a 1-
3 day forecast model (short range forecast) of northern Thailand. For predicting daily
temperature, it is used the same method as weather forecasts. Meteorologists, or forecasters,
use several branches of meteorological knowledge, as well as their experiences, in deciding
for the forecast. Moreover, various climate models help support decision making in prediction
weather alteration.
Developed by using artificial neural networks (ANN), this kind of model learns
patterns of various meteorological events that took places in the past. They have been then
applied to temperature prediction. This learning is compared to experience of weatherman
who likely forecasts more accurate temperature after learning more and more complex forms.
ANN's feature is the well prediction of variable having non-linear relation [3.[
The topography of the north is highlands with long mountains adjoining in large
numbers along with north-south direction. It is the source of many major rivers that flow
together in the region. The cordillera has an average elevation of about 1600 meters above sea
level. It is cold in winter and hot in summer [1.[
Based on information about Thailand's topography, the north is an area of lowest
temperatures in cold season and hottest in summer. With this nature of dominant distribution
of temperature, the northern region is then interesting for this field of temperature study. From

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
statistics of temperature measurement, the highest temperature of 44.5 degrees Celsius was
found in Uttaradit province (27 April 1960), and the lowest temperature was 0.8 degrees
Celsius at Tak (27 Dec 1999). [2]
In Iran, Hayati and Mohebi have studied the short-range prediction of temperature (12
hours) by ANN techniques with multi layers. The data used were during 1996-2006. It was
found that very few errors (MAE) were in range of 0.0079 to 1.2916. This showed that the
developed model was suitable for short-term temperature forecast for Iran [4]. Smith et al. have
developed a neural network model for predicting temperature to mitigate the impact of
agriculture due to the cold weather. It was found that the data input amount of iterations and
the node were increased. Then the MAE reduced for 12-hour temperature forecast. Moreover,
the MAE value had most decreased at 12.5% for 4-hour prediction [5].
In 2013 Kirtsaeng et al. have developed a model to predict daily temperature for
Bangkok, Thailand, using Artificial Neural Network which used variables containing daily
temperature (T), the relative humidity (RH), the pressure (P) and the date order of the year
(Julian date). As a result, the performance of the model, which was based on the R-square
value of temperatures of both model and measurement, was in range of 0.62 to 0.66 and 0.55
to 0 .5 9 for the daily highest temperature and daily lowest temperature, respectively. This
showed that the lowest emperature model performed better the highest temperature model
and that an increasing in number of baptism (Nodes) improved the model efficiency [6]. Later
in 2014, Kirtsaeng et al. have developed a model to predict daily temperature of 24 hours, 48
hours and 7 2 hours at Don Mueang Airport using meteorological elements of Bangkok
weather stations since 1981-20 12, including the 20 13 data used to evaluate its performance.
Comparing with daily temperature measurement, the model performed the R-squad of highest
and lowest temperatures in range of 0 :43 to 0:5 7 and from 0 .61 to 0.7 8 , respectively.
Moreover, the 2 4 -hr forecast performed better than 4 8 -hour and 7 2 -hr expectation,
respectively [7].
This study used data imported, meteorological elements observed by the Thai
Meteorological Department (TMD). These data were, based on international standards and
modern, exchanged with members of the World Meteorological Organization (WMO). Hence,
it made the model process immediately after ending measurement. In comparison with current
dynamic model, ANN advantages were fewer amounts of computing resources and quicker
processing. Expected to covers northern region, this research has been conducted developing
a model to predict daily temperature daily with ANN using meteorological data (temperature,
pressure, humidity) from northern meteorological stations during 1980-2012.
Methodology
Data used
To begin the model process, import the data from 24 northern meteorological stations
during 10 June 1980 to 27 October 2011 )total 11,555 days). The data covers the daily
minimum temperatures, the maximum temperature, the relative humidity, the air pressure and
the order of dates of years, at 00UTC (07LST) for forecasting the lowest temperature and at
09UTC (16LST) for predicting the highest temperature [6.[
The model has been validated by the 239-day data series during 2014-2015 . There are

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
6 series consisting of 24-hr Tmax, 48-hr Tmax, 72-hr Tmax, 24-hr Tmin, 48-hr Tmin and 72-hr
Tmin.
Figure 1: Meteorological stations of Northern Thailand
Model setup
This model uses 3-Layer Back Propagation Algorithm by determining 20 nodes and
5,000interaction [7 [ to create daily temperatures of 24-, 48- and 72-hour ahead.
(A) Maximum temperature forecast (B) Minimum temperature forecast
Figure 2: Schematic diagram of modeling Artificial Neural Networks
Evaluation method
Accuracy of developed model was assessed by mean of Mean Absolute Error (MAE)
and Root Mean Squared Error (RMSE) [8].



n
kkk oy
n
MAE
1
1
(1)

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |




n
kkk
oy
n
RMSE
1
2
1
(2)
where y is the forecast value and "o" equals to measurement value.
Results and Discussions
Of the 12 ANN models, which used 3-Layer Back Propagation Algorithm by defining
20 nodes and 5, 000 interactions, their efficiency showed the R-squared value as Figure 3
shown. The models learned and remembered events from imported data that were from 24
northern meteorological stations since 10 June 1980-27 October 2011 )total 11, 555 days).
About maximum temperature forecast ranging 24, 48 and 72 hours, the model released an
average R-squared 0.707 , 0.565 and 0.505 while for minimum temperature of 24, 48 and 72
hours, it had 0.933 , 0.812 and 0.777 , respectively. The minimum temperature was performed
better than the maximum temperature.
Figure 3: Example of Scatter plot (24 hour forecast) for Mae Hong Son Station
The below figures 4-5 showed MAE and RMSE of learning daily temperature. The
MAE of maximum temperature forecast ranging 24, 48 and 72 hours, granted following
ranges: 1.03-1.38, 1.27-1.72 and 1.18-1.84, with an average of 1.20, 1.50 and 1.60. For minimum
temperature of 24, 48 and 72 hours, the MAE gave ranges: 0.85-1.12, 1.02-1.35 and 1.10-1.46 .
The average MAE had 0.98 1.20 and 1.30 , respectively.
The RMSE of maximum temperature forecast ranging 24, 48 and 72 hours, granted
following ranges: 1.38-1.87, 1.66-2.27 and 1.54-2.42, with an average of 1.63, 1.98 and 2.11. For
minimum temperature of 24, 48 and 72 hours, the RMSE gave ranges: 1.15-1.55 , 1.40-1.84 and
1.50-1.99 with an average of 1.32 , 1.
60 and 1.73 , respectively.

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
Figure 4: The MAE, RMSE, R
2
of maximum temperature prediction
Figure 5: The MAE, RMSE, R
2
of minimum temperature prediction
Model verification
The model efficiency of daily temperature forecast is shown as Tab le 1. The MAE of
maximum temperature forecast ranging 24, 48 and 72 hours, granted following ranges: 1.00-
1.81 , 1.22-1.80 and 1.30-1.99 , with an average of 1.32, 1.54 and 1.67 . For minimum
temperature of 24, 48 and 72 hours, the MAE gave ranges: 0.90-1.19 , 1.06-1.47 and 1.14-1.80
with an average of 1.04 , 1.23 and 1.35 , respectively.
The RMSE of maximum temperature forecast ranging 24, 48 and 72 hours, granted
following ranges: 1.28-1.89 , 1.54-2.24 and 1.63-2.42 , with an average of 1.63 , 1.93 and 2.05 . For
minimum temperature of 24, 48 and 72 hours, the RMSE gave ranges: 1.15-1.62 , 1.30-1.94 and

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
1.40-2.15 , with an average of 1.37 , 1.61 and 1.76 , respectively.
It was found that the model provide short-period temperature forecast more accurate
than long-term temperature. Therefore, the developed model could predict minimum
temperature more efficient than highest temperature.
Table 1. The error value of minimum and maximum temperature in northern Thailand (cont.)

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
Conclusion
For northern region of Thailand, the daily temperature forecast by ANN model found
it performed well with 1-3 days temperature prediction. In comparison, the lowest temperature
was slightly more accurate than maximum temperature. For short-term forecast, it gave a
better prediction than longer term forecast, consistent with results of previous studies. [7]
In this study, pointed value forecasting would be analyzed as spatial prediction
outcome. So we knew where the temperature was high or low that it was able to be harmful to
health or it might hinder human beings activities.
For further use in future studies, temperature forecast should be analyzed in the spatial
form. This allowed us seeing its distribution and the center of region where the temperature
was high or low, as well as the prediction of the areas where there was no monitoring stations.
References
[1] S. Kirtsaeng and S. Kirtsaeng,"Analysis and simulation of heat index for developing a
heat alert system over Thailand",The 2015 Asian Conference on Defence
Technology, 2015, 63-68.
[2] Thai Meteorological Department (www.tmd.go.th/info/knowledge_weather01_ n.html) ,
2559
[3] P. Viotti, G. Liuti and P.D. Genova, "Atmospheric urban pollution: applications of an
artificial neural network (ANN) to the city of Perugia."Ecological Modelling, Vol. 148,
no. 1 (2012): 27-46
[4] Mohsen Hayati and Zahra Mohebi, "Application of Artificial Neural Network for
Temperature Forecasting "International Journal of Electrical, Computer,

| International Conference on Disaster Management: From Polar Region to the Local Communities Social and Environmental Development National
Institute of Development Administration (NIDA)|

ARTICLE IN PRESS
| Social and Environmental Development National Institute of Development Administration (NIDA)
International Conference Disaster Management: From Polar Region to the Local Communities |
Energetic, Electronic and Communication Engineering, Vol:1, No:4, 2007, 654-
658
[5] Brian A. Smith, Ronald W. McClendon, and Gerrit Hoogenboom. "Improving air
temperature prediction with Artificial Neural Networks"International Journal of
Computer, Electrical, Automation, Control and Information Engineering, Vol:1,
No:10, 2007, 3146-3153
[6] S. Kirtsaeng, S. Kirtsaeng, P. Sukthawee and F. Masthawee, "Improvement of Daily
Temperature Prediction Model for Bangkok using Artifical Neural Networks", 39
th
Congress on Science and Technology of Thailand, 2013, 5pgs.
[7] S. Kirtsaeng, S. Kirtsaeng and P. Sukthawee, "Application of Artifical Neural Networks
for Daily Temperature Prediction over Don Mueang International Airport", 1
st
Environment and Natural Resources International Conference, 2014, 67-71.
[8] D.S. Wilks,"Statistical Methods in the Atmospheric Sciences."2th ed. International
Geophysics, 2006. p. 278-279.
[9] N.S.N. Lam, "Spatial Interpolation Methods: A Review" The American
Cartographer,Vol:10, No:2, 1983, 129-149
[10] C.A. Karavitis, S.Alexandris, D.E. Tsesmelis and G. Athanasopoulos, "Application of
the Standardized PrecipitationIndex(SPI)in Greece." Water,2011, 3, 787-805