Professional Documents
Culture Documents
1757
G. HOOGENBOOM
Department of Biological and Agricultural Engineering, University of Georgia, Griffin, Georgia
R. W. MCCLENDON
Artificial Intelligence Center, and the Department of Biological and Agricultural Engineering, Driftmier Engineering Center,
University of Georgia, Athens, Georgia
ABSTRACT
Dewpoint temperature, the temperature at which water vapor in the air will condense into liquid, can be
useful in estimating frost, fog, snow, dew, evapotranspiration, and other meteorological variables. The goal
of this study was to use artificial neural networks (ANNs) to predict dewpoint temperature from 1 to 12 h
ahead using prior weather data as inputs. This study explores using three-layer backpropagation ANNs and
weather data combined for three years from 20 locations in Georgia, United States, to develop general
models for dewpoint temperature prediction anywhere within Georgia. Specific objectives included the
selection of the important weather-related inputs, the setting of ANN parameters, and the selection of the
duration of prior input data. An iterative search found that, in addition to dewpoint temperature, important
weather-related ANN inputs included relative humidity, solar radiation, air temperature, wind speed, and
vapor pressure. Experiments also showed that the best models included 60 nodes in the ANN hidden layer,
a ⫾0.15 initial range for the ANN weights, a 0.35 ANN learning rate, and a duration of prior weather-
related data used as inputs ranging from 6 to 30 h based on the lead time. The evaluation of the final models
with weather data from 20 separate locations and for a different year showed that the 1-, 4-, 8-, and 12-h
predictions had mean absolute errors (MAEs) of 0.550°, 1.234°, 1.799°, and 2.280°C, respectively. These
final models predicted dewpoint temperature adequately using previously unseen weather data, including
difficult freeze and heat stress extremes. These predictions are useful for decisions in agriculture because
dewpoint temperature along with air temperature affects the intensity of freezes and heat waves, which can
damage crops, equipment, and structures and can cause injury or death to animals and humans.
DOI: 10.1175/2007JAMC1693.1
calculate critical-damage air temperature for specific ences. Gardner and Dorling (1998) review ANNs used
crops, allowing producers to determine the potential for prediction of ozone concentration and daily maxi-
for a frost to damage those crops (Snyder and de Melo- mum ozone, sulfur dioxide concentration, tornadoes,
Abreu 2005). Heat waves, which cause damage and thunderstorms, solar radiation, carbon dioxide, pollut-
take the lives of people and animals, are intensified by ants and monsoon rainfall. More recently, Bodri and
high dewpoint temperatures (Sandstrom et al. 2004). Cermak (2000) developed an ANN using 38 yr of rain-
These studies suggest that a non-location-specific dew- fall data to predict monthly and yearly precipitation
point temperature prediction would be helpful for levels for multiple sites in the Czech Republic. Using
those in rural areas where fewer weather predictions spatial and temporal data of recent rainfall, Luk et al.
are made, and there are more crops and livestock. A (2000) developed an ANN for short-term precipitation
study by Robinson (2000) proposed that the dewpoint prediction focused on predicting flash flood rainfall
temperature in the United States is slowly increasing amounts for 15 min ahead for various areas of western
over time and, therefore, could be an important Sydney, Australia. Maqsood et al. (2004) used an en-
weather variable for studies on long-term climate change. semble of ANNs to provide 24-h predictions for air
Hubbard et al. (2003) developed a regression model temperature, wind speed, and humidity at the Regina
for estimating the daily average dewpoint temperature, Airport in Canada. Wedge et al. (2005) developed an
using the daily mean, minimum, and maximum air tem- ANN for prediction of waves spilling over sea walls in
perature as inputs. Their research used 14 yr of data for using sea conditions and wall properties as inputs.
six cities from South Dakota, Nebraska, Colorado, and Steidley et al. (2005) use ANNs to predict tidal water
Kansas in the United States. Their regression equation levels for periods of 3–48 h ahead for a shallow embay-
based on multiple cities was more accurate than the ment on the coast of Texas in the United States.
regression equations for each of the individual cities, Jain et al. (2003, 2006) developed ANNs to predict
with a mean absolute error (MAE) of 2.2°C for the hourly air temperatures for 1–12 h for three locations in
most accurate regression equation. This study’s estima- Georgia, United States, using inputs of current air tem-
tions are useful for determining the values for missing perature, relative humidity, solar radiation, and wind
historical weather data, but did not allow the prediction speed, and up to 6 h of prior data. The MAEs for each
of future values. location varied from 0.6° to 0.7°C for the 1-h lead time
Diab and Saade (1999) used a fuzzy inference system and from 2.4° to 3.0°C for the 12-h lead time. Smith et
with rules developed based on their own intuition about al. (2006) improved on the results of Jain et al. (2003,
the correlation of weather variables to predict dewpoint 2006) by using cyclic variables to represent the day of
temperature for exactly 24 h ahead. The inference rules year and time of day as additional inputs to the ANN.
used the season of the year, barometric pressure, air Smith et al. (2006) also trained multiple ANNs with the
temperature, and wind speed as inputs, each with their same parameters but different initial weights and found
own fuzzy membership functions while the output dew- that the minimum error on multiple networks provided
point temperature membership functions were ex- an improved comparison during model development,
pressed as low, medium, and high. The evaluation with with an MAE of 0.54°C for a 1-h prediction and 2.33°C
40 uniformly distributed days for all four seasons in for a 12-h prediction for the evaluation dataset.
1994 resulted in absolute errors ranging to a maximum Mittal and Zhang (2003) developed ANNs to esti-
of 8°C with no mean error presented. mate several weather variables using other weather
An artificial neural network (ANN) is a robust com- variables, a process used for estimating missing histori-
putational technique modeled after biological neuron cal data. Their ANN estimations provided an alterna-
connections found in human brains (Bose and Liang tive to the traditional estimations done with psychro-
1995; Haykin 1999). Like the human brain, ANNs are metric charts and mathematical models. They devel-
repeatedly exposed to inputs and vary the strength of oped an ANN model to estimate wet-bulb temperature,
the connections between neurons based on those in- enthalpy, humidity ratio, specific volume, and dewpoint
puts. Thus learning for most ANNs is accomplished temperature using dry-bulb temperature and relative
using an iterative process instead of single calculation humidity as inputs. Their dataset included values ob-
as would be used with most types of regression and tained from the psychrometric charts that did not cor-
Bayesian classification. ANNs have been used to help respond to actual historical data or specific locations,
solve many real-world problems such as pattern match- but rather to known relationships among weather vari-
ing, classification, and prediction (Bose and Liang 1995; ables. The MAE for the dewpoint temperature estima-
Gardner and Dorling 1998; Haykin 1999). tion was 0.305°C.
Often ANNs have been used in the atmospheric sci- Dewpoint temperature has been estimated (Kimball
FIG. 1. AEMN weather stations; 20 sites were selected for model development and 20 sites
were selected for model evaluation.
et al. 1997; Mahmood and Hubbard 2005; Mittal and cision making in agricultural production and natural
Zhang 2003; Parlange and Katz 2000) and analyzed for resource management (additional information is avail-
long-term trends (Robinson 1998, 2000; Sandstrom et able online at http://www.georgiaweather.net; Hoogen-
al. 2004), but there is little research on short-term dew- boom 2000). With over 70 weather stations distributed
point temperature prediction. The overall goal of the throughout Georgia, each station measures weather
research presented herein was to develop ANN models data for variables including air temperature, relative
for predicting hourly dewpoint temperatures for up to humidity, vapor pressure, wind speed and wind direc-
12 h in advance. Specific objectives were to identify the tion, solar radiation, atmospheric pressure, and rainfall.
important weather-related inputs that affect dewpoint Vapor pressure deficit and dewpoint temperature are
temperature prediction, to determine the preferred val- calculated based from these variables using standard
ues of the ANN parameters, and to determine the pre- methods. The total amount of rainfall and average of
ferred duration of prior data for each lead time. The every other variable is determined for each 15-min in-
application of the resulting models would be to predict terval based on 1-s observations. Although the AEMN
short-term hourly dewpoint temperature for any loca- data collection began for some locations in 1992, the
tion in Georgia. determination of dewpoint temperature only started in
September 2002 following requests from the horticul-
2. Methodology tural industry.
Data from 40 of the AEMN weather stations were
a. Data used in this study: 20 sites were used for model devel-
The University of Georgia Automated Environmen- opment, and 20 different sites were used for model
tal Monitoring Network (AEMN) provides Web-based evaluation. These sites were selected to represent the
delivery of current and historical weather data, as well geographic and regional diversity of Georgia as shown
as weather-based tools and applications useful for de- in Fig. 1. The years 2002–04 were used for model de-
冦
1
tanh共n兲 for 0⬍jⱕ J
3
2 1 2
fj共n兲 ⫽ e⫺n for J⬍jⱕ J, 共4兲
3 3
2 2
1 ⫺ e⫺n for J⬍jⱕJ
3
where n is the input to the activation function.
The second mode of the ANN is backpropagating the
error to adjust the weights. The weight adjustment
⌬(j) for each weight from the hidden layer to the out-
put node j is defined as
J
冉
⌬共␣ji兲 ⫽ f ⬘j ␣j0 ⫹ 兺
I
i⫽1
冊冉
␣jixi g⬘ 0 ⫹
J
兺y
j⫽1
j j 冊
⫻ 共t ⫺ z兲j xi, 共6兲
FIG. 3. Ward EBP ANN architecture with a single hidden layer where is the learning rate and t is the target output
consisting of three slabs of hidden nodes with different activation value. The nodes y0 and x0 are bias nodes that are al-
functions: Gaussian, Gaussian complement, and hyperbolic tan-
gent.
ways set to 1, although their corresponding weights,
0 and ␣j0, are adjusted. The adjustments for all the
weights were applied after each training pattern and
work by Jain et al. (2003) and Smith et al. (2006), who this is referred to as a learning event. An EBP ANN
showed that this type of ANN was suitable for air tem- with all the parameters including inputs, initial weight
perature prediction. range, number of hidden nodes per slab, and learning
An EBP ANN model has two modes. The first is a rate is referred to herein as a model. A single instan-
feed-forward mode where a set of inputs xi, where i tiation of the model with random initial weights and a
ranges from 1 to I, is mapped to a single output z by the randomly ordered set of training observations selected
following equations: from the development dataset is referred to as a net-
冉 兺y 冊
J work.
z ⫽ g 0 ⫹ j j and 共1兲 Traditionally, EBP ANNs use observations in a train-
j⫽1 ing dataset to search iteratively for an optimal set of
I weights that connect the nodes between adjacent lay-
yj ⫽ fj 共␣j0 ⫹ 兺 ␣ x 兲,
i⫽1
ji i 共2兲 ers. A separate testing dataset is used to stop the train-
ing when the testing dataset error reaches a minimum.
where ␣ji are the weights from the input layer to the A separate selection dataset is used as a dataset to
hidden layer, j are the weights from the hidden layer judge the error of that network after training has been
to the output node, and yj is the output of the nodes in stopped and a comparison of selection dataset errors is
the hidden layer, where j ranged from 1 to J. The lo- used for selection of parameters during model devel-
gistic activation function g is defined as follows: opment. Preliminary tests indicated that a testing
dataset was not necessary if the training dataset was
1 sufficiently large. The mean absolute error of the train-
g共n兲 ⫽ , 共3兲
1 ⫹ e⫺n ing and testing datasets, though slightly different, al-
ways tracked each other with as few as 20 000 indepen- work set, would be used to determine the accuracy of a
dent observations in each dataset (e.g., as the training model during model development. The population then
dataset error continued to decrease, the testing dataset would be all instances of a model, and each network
error did as well). would be one sampling. As with any distribution, the
Overtraining, where the ANN is able to make accu- more samples the higher the statistical validity, but
rate predictions on the training dataset, but not on the computational time was also a consideration. One net-
testing and selection datasets can be a problem for work trained for 1 000 000 learning events required 1–6
ANNs. These preliminary tests indicated overtraining h of computational time depending on the parameters,
was not a concern as the error continued to decrease on making 30 networks require 30–180 h of computational
both training and testing datasets even after 5 000 000 time. Once trained, a network evaluated in feed-
learning events when the training and testing datasets forward mode on 100 000 observations required only
consisted of at least 100 000 observations. This held several minutes. All tests were conducted on 36 per-
true for networks with a single or multiple weather- sonal computers, that is, 32 Pentium 4 and 4 Pentium 3
related inputs. It was, therefore, decided that a testing computers, in the computer laboratories of the Depart-
dataset would not be used and a fixed number of learn- ment of Biological and Agricultural Engineering at the
ing events would be used to stop training. An epoch is University of Georgia.
one pass through all the observations in the training Because of the nonnormal distribution of the errors
dataset. With 100 000 observations in the training for a network set, a number of different statistical mea-
dataset the decrease in the mean absolute error from surements were considered to approximate the error of
epoch 9 to 10 was typically less than 0.01°C and always a model based on a network set. A preliminary test
less than 0.06°C, so the stopping criteria for training considered seven statistical measurements. Four statis-
was arbitrarily fixed at 10 epochs of 100 000 observa- tical measurements of central tendency were consid-
tions each, that is, 1 000 000 learning events. For model ered: the mean, the mean truncated 20%, the mean
development, the selection dataset errors were com- truncated 40%, and the median. A truncated or
pared in order to select values of ANN parameters for trimmed mean is the mean of the remaining values after
determining the most accurate model. a percentage is removed, half from the minimums and
All model development was conducted using data half from the maximums. The truncated mean is useful
from the development dataset, which consisted of ap- as a robust measure of central tendency, especially for
proximately 1 560 000 observations. Because the devel- asymmetric distributions (Marazzi and Ruffieux 1999).
opment data spanned less than three years, it was not Three minima were considered: the minimum, the av-
partitioned into separate years for the training and se- erage of the minimum 5, and the average of the mini-
lection datasets. To obtain the best representative mum 10. Of all seven statistical measures, the average
sample the training and selection datasets consisted of of the minimum five provided the smallest range and
100 000 observations randomly selected without re- standard deviation among the 30 network set instantia-
placement from the development dataset for each net- tions for the preliminary test, suggesting this statistical
work. The training and selection datasets were inde- measure may be more stable than the minimum which
pendent, and each represented 20 cities for three years had been used in ANN temperature prediction research
of data from 2002 to 2004. (Jain et al. 2003; Smith et al. 2006). Therefore, the av-
erage of the minimum five MAEs for a network set was
c. Measurement of model error used to approximate the error of a model and was re-
The MAE between predicted and observed dewpoint ferred to as the MAE for that model during model
temperature for a particular dataset was selected as the development.
measure of accuracy. For model development this was
the selection dataset, and for model evaluation this was
d. Experimental procedure
the evaluation dataset. Because each network was in- In the first experiment, the preferred set of weather-
stantiated with random weights and the training and related inputs was determined using a 6-h lead time and
selection datasets were selected and ordered randomly, an 18-h duration of prior data. Dewpoint temperature,
networks representing the same model produced differ- air temperature, relative humidity, vapor pressure, va-
ent MAEs. Thirty observations were considered as an por pressure deficit, wind speed, solar radiation, and
adequate population sample to closely approximate sta- rainfall were the inputs considered in this search. Dew-
tistical measurements such as the mean for the popula- point temperature was selected as the first weather-
tion (Freund and Wilson 1993). Therefore, it was arbi- related input considered, and the remaining predictors
trarily decided that 30 networks, referred to as a net- were selected in an iterative fashion: ANN models us-
TABLE 1. The effect of selected weather-related input combinations on dewpoint temperature prediction for the development
dataset using artificial neural networks.
* Average of the minimum five MAEs of the selection dataset out of 30 networks.
** The best variables selected based on the minimum MAE.
ing dewpoint temperature paired with each of the re- the following ANN parameters were determined: num-
maining weather-related inputs were compared, and ber of hidden nodes per slab, initial weight range, and
the input that produced the lowest error was retained. learning rate. As each parameter was varied the current
The two weather-related inputs were then combined preferred model was determined by the model with the
with each of the remaining inputs to form models with minimum error. This experiment used the previously
three weather-related inputs, and the one with the low- determined weather-related inputs and a 6-h lead time.
est error was retained. This process was continued until In the third experiment the preferred duration of
all the possible inputs were included or models with prior data was determined for each lead time and var-
additional inputs did not produce a smaller error than ied between 1 and 12 h. Preliminary tests indicated that
the previous preferred model. The initial ANN param- the duration of prior data was correlated to the lead
eters were arbitrarily chosen to be 20 hidden nodes per time and a search for the preferred model for all lead
slab for a total of 60 nodes in the hidden layer, a learn- times should range from 6 to 30 h. To help ensure the
ing rate of 0.1, and an initial weight of ⫾0.2. reliability of the preferred model for each lead time,
In the second experiment, the preferred values for two models with longer durations of prior data and two
TABLE 2. The effect of the duration of prior data and lead time on dewpoint temperature prediction based on the MAE* and its CV.
* Average of the minimum five MAEs of the selection dataset for 30 networks.
** Duration of prior data selected for each lead time.
models with shorter durations of prior data were also ANN with six weather-related inputs resulted in the
developed. lowest MAE, 1.463°C (Table 1). Vapor pressure deficit
The training for the final evaluation was conducted and amount of rain did not improve model accuracy
by training 30 networks for each of the 12 lead times when they were included.
using half of the development dataset (i.e., approxi- For the second of the experiments described in sec-
mately 780 000 observations) for a training dataset in- tion 2d, the number of hidden nodes per slab was varied
stead of the 100 000 observations used during model from 10 to 70 in increments of 10. The MAE decreased
development. The other half was used as a selection from 1.471°C to 1.463°C when the number of hidden
dataset to choose the preferred network for each lead nodes was increased from 10 to 20 nodes per slab, but
time model. The 12 preferred networks, which repre- thereafter increasing the number of nodes per slab had
sented the 12 final models, were used in feed-forward a negligible effect on accuracy. Therefore, the number
mode for model evaluation on the evaluation dataset. of hidden nodes per slab selected was 20.
The predictions of the 12 final models were directly The range of initial weights was varied from ⫾0.05 to
compared with predictions using the current dewpoint ⫾0.40 in increments of 0.05. An initial weight range of
temperature as the predicted value. ⫾0.15 resulted in the lowest MAE of 1.463°C and was
therefore selected for further model development. The
learning rate was varied from 0.05 to 0.60 in increments
3. Model development
of 0.05. A model with a learning rate of 0.35 had the
As stated in the previous section during the search lowest MAE, 1.445°C, and was selected for further
for the important weather-related inputs, several values model development.
were held constant including 20 hidden nodes per slab, For the third of the experiments described in section
an 18-h duration of prior data, a 0.1 learning rate, a 0.2 2d, the duration of prior data was varied from 6 to 30 h
initial weight range and a lead time of 6 h. The fuzzy in increments of 6 h for the 12 lead times, and in some
membership function inputs for time of day and day of cases the range of the duration was extended (Table 2).
year (Fig. 2) were also included in each model that was The model with the lowest MAE and, in the case of a
developed. When dewpoint temperature was consid- tie, the lowest coefficient of variation (CV) for the
ered as the only weather-related input, the MAE was MAE, for each lead time was selected as the best model
1.620°C (Table 1). The dewpoint temperature–only for model development. The best models for the 1- and
model was then coupled with each possible remaining 2-h lead times were the models that included 6 h of
weather-related input to determine the best two- prior data. The best models for the 5-, 6-, 7-, 9-, and
weather-variable ANN. The model with dewpoint tem- 12-h lead times included 18 h of prior data. The best
perature and relative humidity produced the lowest models for the 3-, 4-, 10-, and 11-h lead times included
MAE of 1.521°C. Continuing with this approach, the 24 h of prior data. The best model for the 8-h lead time
weather-related inputs in order of importance with re- included 30 h of prior data. Therefore, these values did
spect to weather variables 3–6 were solar radiation, air not show a clear relationship between duration of prior
temperature, wind speed, and vapor pressure. The data and lead time.
TABLE 2. (Extended)
FIG. 4. Performance of predicted dewpoint temperature for the evaluation dataset for the (a) 1-, (b)
4-, (c) 8-, and (d) 12-h prediction models.
TABLE 3. Comparison of the MAEs of using the current dewpoint temperature as the predicted dewpoint temperature and using
the final ANN model for the evaluation dataset for lead times of 1–12 h.
* Column values for improvement may not subtract exactly because of rounding.
model evaluation with predictions using the current 0.3°C and 0.4°C for the 4–12-h models. The percent
dewpoint temperature as the predicted temperature for improvement for the 2–10-h models was relatively simi-
the same observations (Table 3). The improvement of lar, ranging from 15.5% to 21.4%, but differed at the
the ANN model over the current dewpoint temperature lower end, where the 1-h model improved by only
was 0.035°C for the 1-h model, 0.162°C for the 2-h 6.0%, and at the higher end where the 11-h model im-
model, 0.212°C for the 3-h model, and varied between provement was 14.7% and the 12-h model improve-
FIG. 5. Predicted dewpoint temperature for freezing conditions for Dahlonega, GA, with (a) 1-, (b) 4-, (c) 8-, and (d) 12-h
prediction models in March 2005. Observed dewpoint temperature and air temperature are also shown.
FIG. 6. Predicted dewpoint temperature for extreme heat condition for Statesboro, GA, with (a) 1-, (b) 4-, (c) 8-, and (d) 12-h
prediction models in August 2005. Observed dewpoint temperature and air temperature are also shown.
ment was 11.9%. The poorer prediction ability for the tremes of low and high air temperature conditions to
lower- and higher-hour models suggests that develop- demonstrate the prediction of dewpoint temperature
ment of the preferred ANN models for those lead times for these situations. A sample period from 18 to 20
might require searching for the optimal model param- March 2005, for Dahlonega, Georgia, was selected as
eters for those lead times separately. The development an example of two early morning freezes in late winter.
of these models included an iterative search for the This scenario would represent a situation in which a
preferred duration of prior data for each lead time, but fruit crop could experience catastrophic damage from
did not include the search for preferred weather-related frost during the blooming phase of the crop. The pre-
inputs and the ANN model parameters for each lead dictions of dewpoint temperature for the 1-, 4-, 8-, and
time. Because those searches were only conducted for 12-h models indicated more accurate predictions for the
the 6-h lead time, the models may be biased in favor of shorter than the longer lead times during these winter
6-h predictions and not generalize as well to the lower freezes (Fig. 5). The 1- and 4-h predictions showed a
and higher lead times. A full comparison with other drop in dewpoint temperature during the freeze event,
methods of dewpoint temperature prediction was out- but the 4-h model placed it later than it actually oc-
side the scope of this paper, but would be an appropri- curred. In contrast, the 8- and 12-h models did not pre-
ate follow-up study to understand how ANN predic- dict the drop in dewpoint temperature, but instead pre-
tions compare to other intelligent methods, such as ad- dicted that it would remain steady around 1°–3°C. Simi-
vanced regression models and Bayesian learning, and larly, the low dewpoint temperature between 1200 and
how other intelligent methods compare to physical 1800 h on 19 March was predicted well by the 1-h pre-
weather models. diction, fairly accurately with the 4-h model predicting
Sample periods were selected that included the ex- an even lower dewpoint temperature, and less accurate-