Sie sind auf Seite 1von 9

On estimating soil surface temperature proles

G. Mihalakakou
*
Laboratory of Meteorology, Division of Applied Physics, Department of Physics, University of Athens,
University Campus, Building PHYS-V, Athens GR 15784, Greece
Received 10 March 2001; accepted 15 May 2001
Abstract
The present study deals with two methods for modeling and estimating the daily and annual variation of soil surface temperature. Soil
surface temperature is an important factor for calculating the thermal performance of buildings in direct contact with the soil as well as for
predicting the efciency of earth-to-air heat exchangers. The two estimation methods are a deterministic model and a neural network
approach. The two methods are tested and validated against extensive sets of measurements for bare and short-grass covered soil in Athens
and Dublin. Finally, the comparison of the two models showed that the proposed intelligent technique is able to adequately estimate the soil
surface temperature distribution. # 2002 Elsevier Science B.V. All rights reserved.
Keywords: Soil surface temperature; Heat conduction equation; Energy balance equation; Data-driven analysis
1. Introduction
Soil temperature is an important parameter in solar energy
applications such as the passive heating and cooling of
buildings and agricultural greenhouses. These applications
can be developed by the building's direct earth contact, which
involves partial or total placing of the building envelope in
direct contact with the soil [1], or by the building's indirect
contact, which involves the use of a buried pipe through
which air from indoors or outdoors of the building is circu-
lated and then brought into the building or the agricultural
greenhouse [24].
The use of direct or indirect earth-coupling techniques for
buildings requires knowledge of the ground temperature
distribution especially at surface. It is still commonly
believed that soil temperature distribution at any depth below
earth's surface remains unchanged throughout the year.
However, various researchers showed that soil temperatures
at shallow depths present signicant uctuation on both
daily and annual basis [57]. According to their studies
the heat ow inside the earth is inuenced by several para-
meters, such as solar radiation, air temperature, wind speed,
time of the year, shading, soil properties, etc. which present a
seasonal or irregular variation. For this reason, prediction and
estimation of soil temperature is rather difcult, especially
near the ground surface where the soil temperature variations
are the highest.
Measurements of soil temperature at surface and at
various depths are spatially and temporally limited. In
locations where soil temperature measurements are sparse,
theoretical estimates can be used to predict it from other
existing data, which depend strongly on the local meteo-
climatic conditions and soil properties.
Using the existing data, various mathematical models for
the prediction of the soil temperature as a function of depth,
season and soil properties have already been developed.
These models are based on analysis of multiyear measure-
ments [8,9] or on Fourier analysis [10,11]. The development
and application of such models facilitate the estimation
procedure as they provide a continuous spectrum of values
while at the same time they can provide information on
parameters which are not measured directly, like the soil
temperature at higher depths or the diurnal variation of the
soil surface temperature. Although these models describe
the temperature prole at various depths below ground
surface, they do not take into account the weather conditions
described by several climatic parameters.
An analytical model, based on the transient heat conduc-
tion differential equation as well as on the energy balance
equation at the ground surface, is used in the present study to
estimate the soil surface temperature. The energy balance
equation involves energy exchange between air and soil, the
solar radiation absorbed by the soil surface as well as the
long-wave radiation. The above model has been selected
because although it is based on heat conduction differential
equation, it also takes into account the prevailing weather
conditions and soil properties at any location. The model is
Energy and Buildings 34 (2002) 251259
*
Tel.: 30-1-727-6850; fax: 30-1-729-5282.
E-mail address: pmihalak@cc.uoa.gr (G. Mihalakakou).
0378-7788/02/$ see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S0 3 7 8 - 7 7 8 8 ( 0 1 ) 0 0 0 8 9 - 5
validated using extensive sets of measurements for bare and
short-grass soil in Athens and Dublin, and a close agreement
between the calculated and measured values of soil surface
temperature is observed. Its main disadvantage is that it
requires several input parameters which very often are not
available, as the climatic measurements network in devel-
oped countries is still in progress and there are many
locations where measured data are rather sparse.
Moreover, a neural network approach is used in the
present study for estimating the soil surface temperature
using as inputs several climatic parameters. Neural network
models belong to the class of data-driven approaches instead
of model-driven approaches [12]. In the data-driven models,
the analysis depends only on the available data, with little
rationalization about possible interactions. Relationships
between variables, models, laws, and predictions are con-
structed after building a machine that simulates the con-
sidered data. Neural network systems have been shown
to present the capability of modeling complex nonlinear
processes to arbitrary degrees of accuracy.
Intelligent data-driven approaches such as neural net-
works present several advantages over conventional, deter-
ministic analytical models. The most important among them
is their simplicity as well as the fact that they do not require
any assumption to be made about the underlying function or
model to be used. All they need are the historical data of
the target and of the relevant input parameters for training
the data-driven system. Once the system is well-trained and
the error between the target and the method estimations has
converged to an acceptable level, it is ready for use. Intel-
ligent data-driven techniques have been designed and used
for several scientic applications [13,14].
The present study aims primarily at presenting and com-
paring two models, one analytical deterministic model based
on the heat conduction and on the energy balance equations
and one intelligent data-driven model, for estimating the
temperature at the soil surface.
The results of the neural network model are compared
with those of the analytic one in order to investigate the
applicability of the intelligent method. The analytic model is
able to give sufciently accurate estimations provided that
all required input parameters are available. This is not
always possible and for this reason the data-driven method
can be very effective as it is simpler and it does not use so
many input parameters as the analytic model.
Secondly, the inuence of the main input climatic para-
meters on the temperature at the soil surface is examined
using the neural network approach, thus giving a quantitative
information on the dependence of soil surface temperature
on weather conditions.
2. Modeling the ground surface temperature
The undisturbed ground temperature can be accurately
described by the following one-dimensional, transient heat
conduction equation [15,16]:
@
2
T(z; t)
@z
2
=
1
a
@T(z; t)@t (1)
where a is the ground diffusivity and z the depth below
ground surface. An analytical solution of Eq. (1) for a semi-
infinite homogeneous solid with constant physical proper-
ties can be reported as follows:
T
sur
(t) = T
m
A
s
Re[exp(iwt)[ (2)
where T
m
is the mean annual ground surface temperature
and A
s
and w are the amplitude and the frequency of the
temperature wave at the ground surface, respectively.
For solving Eq. (2) the following energy balance equation
was used as a boundary condition at the ground surface
[16,17]:

K@T
sur
@y

y=0
= CE LR SR LE (3)
where K is the thermal conductivity of the soil, CE the
convective energy exchanged between air and ground sur-
face, LR the long-wave radiation emitted from the ground
surface, SR the solar radiation absorbed from the ground
surface and LE the latent heat flux due to evaporation.
1. Convective energy: The convective energy (CE), ex-
changed between the air and ground surface is taken
from the following expression [18]:
CE = h
sur
(T
a
T
sur
) (4)
where T
sur
is the ground surface temperature, T
a
the air
temperature above the ground surface and h
sur
is the
convective heat transfer coefficient at the soil surface.
2. Solar radiation: The short-wave solar radiation (SR),
absorbed from the ground surface can be estimated
according to solar radiation analysis from the following
equation [18]:
SR = bS (5)
where b is a coefficient depending on the ground surface
absorptivity and its illumination. For a fully illuminated
ground surface b = 1albedo.
3. Long-wave radiation: The long-wave radiation (LR), for
horizontal surfaces can be considered nearly constant.
Its value can be estimated using the following expres-
sion [19]:
LR = eDR (6)
where e is the emittance of the ground surface and
DR is a term depending on the relative humidity of the
ground and the air above the ground surface, on the
effective sky temperature and on the soil radiative
properties. DR can be calculated using empirical
correlations [20,21].
4. Latent heat flux due to evaporation: The latent heat flux
(LE), from the ground surface caused by evaporation
252 G. Mihalakakou / Energy and Buildings 34 (2002) 251259
can be calculated from the following expression [22]:
LE = 0:0168fh
sur
[(aT
sur
b) r
a
(aT
a
b)[ (7)
where a=103 Pa K
1
, b=609 Pa, 263 K_T _303 K, r
a
is the relative humidity of the air above the ground
surface and f is a fraction, which depends mainly on the
ground cover and on the humidity level of the ground.
The fraction f can be estimated as follows [22]:
For bare soils the fraction f increases with soil humidity
(f = 1 for saturated soils, f = 0:60:8 for moist soils,
f = 0:40:5 for dry soils and f = 0:10:2 for arid
soils).
For grass covered soils the fraction f can be calculated
by multiplying the previous values of the fraction f for
bare soils with the coefficient 0.70.
The ground surface temperature T
sur
(t) can be calculated
from Eq. (2) using the energy balance Eq. (3) as boundary
condition at the soil surface. For this purpose the values of
CE, SR, LR and LE, as calculated from Eqs. (4)(7) were
used in Eq. (3) [16].
The accuracy of the above model was tested using two
extensive sets of experimental data for Athens and Dublin.
The National Observatory of Athens (NOA), provided soil
temperature measurements since 1917. The observatory is
situated on a hill in the center of Athens (latitude=37:588N,
longitude = 23:438E and altitude = 107 m). The observa-
tions were performed at the surface of both short-grass
covered and bare soils as well as at depths 0.30, 0.60,
0.90 and 1.20 m below bare soils. The lawn is irrigated
every day at about 09.00 LST during the dry period of the
year. The temperatures at the surface of the bare soil were
taken using a couple of thermometers (normal and extreme),
while the surface temperature for the lawn was measured
using thermometers of minimum and maximum indication.
Since January 1993, temperature values have been recorded
on a continuous basis using data logger system. The thermo-
meters for the short-grass soil surface area are placed
horizontally in the middle of a small area (12 m
2
) of ground
covered by dense lawn. The thermometers are covered
perfectly by the lawn, which has a height of 0.10 m.
Fig. 1. Variation of the measured and predicted with the deterministic model mean monthly ground surface temperatures for short-grass and bare soil, for
Athens and for the 10-year time period (19811990).
G. Mihalakakou / Energy and Buildings 34 (2002) 251259 253
Calculations for the present study extend over a 10-year
time period (19811990) and experimental values were used
for the same period.
Comparison between the ground surface temperature
values predicted by the previously described model and
the measured NOAvalues has shown that the model provides
accurate ground surface temperatures. Fig. 1 shows the
variation of the predicted and measured values of the mean
monthly surface temperature for short-grass covered and
bare soil and for the 10-year time period. As shown, the
measured and predicted values are in close agreement. For
the bare soil the differences are less than 18C, while for the
short-grass covered soil they are less than 1.58C. The root
mean square errors are found equal to 0.28C for bare soil and
0.68C for short-grass soil.
Moreover, the predicted temperature values were com-
pared with ground surface temperature measurements for
Dublin Airport. The Irish Meteorological Service provided
the data for bare and short-grass covered soils. The calcula-
tions cover the time period 19751984. Fig. 2 shows the
predicted and measured values of the mean monthly surface
temperature for Dublin Airport for the 10-year time period
(19751984), and for short-grass covered and bare soil. As it
can be seen the predicted values perform well with the
measured ones. In this case, the root mean square errors
are found equal to 0.38C for bare soil and to 0.28C for short-
grass covered soil.
3. The neural network model
3.1. Neural network architecture
Articial neural networks are computing systems, which
attempt to simulate the structure and function of biological
neurons. How the inter-neuron connections are arranged and
the nature of the connections determine the structure of a
network [23].
The estimation problem using neural network models can
be separated into three successive steps or subproblems:
v Model building or neural network architecture.
v The learning or training procedure.
v The testing or diagnostic checking.
Fig. 2. Variation of the measured and predicted with the deterministic model mean monthly ground surface temperatures for short-grass and bare soil, for
Dublin Airport and for the 10-year time period (19751984).
254 G. Mihalakakou / Energy and Buildings 34 (2002) 251259
In the present research, a multiple layered network based
on backpropagation learning algorithm is designed for the
urban heat island intensity estimation at nighttime for every
location. For this reason, a neural network model is built
separately for each one of the experimental stations. The
selected neural network architecture consists of one hidden
layer of 1523 log-sigmoid neurons followed by an output
layer of one linear neuron. Learning is achieved using the
backpropagation algorithm of Rumelhart et al. [24] to train
the network.
3.2. Results and discussion
The time series generated in this paper was temperature at
the ground surface measured at the NOA station.
Hourly values of air temperature, relative humidity,
global solar radiation, and soil surface temperature for
bare, for 8 years (19911998) and for various months of
the year were used for training and testing the network.
Analytically, 6 years (19911997) were used for training
the neural network and 1 year (1998) for testing the training
data.
The network was trained over a certain part of the climatic
data, and once training was completed, the network was
tested over the remaining data.
The input parameters of the neural network model were
the following:
v air temperature (8C);
v relative humidity measurements (%);
v global solar radiation (MJ m
2
).
The output was the hourly values of ground surface
temperature values for bare soil. Training is performed using
hourly values of the input climatic parameters for the estima-
tion of ground surface temperature for 7 years (19911997)
and for various months of the year. Calculations were
performed for various months of the year, and the following
two time periods were selected for the presentation of results:
1. The cold period, which consists of the months of
November, December, January, February and March.
The month of January was regarded as representative of
the cold period for the results' presentation.
2. The warm period of the year, which consists of the
months of May, June, July, August and September.
Accordingly, the month of July was considered to be the
representative month of the warm period for the
presentation of results.
Fig. 3 presents the comparison of the measured soil surface
temperature values with the neural network estimated once
Fig. 3. Comparison of the measured with the neural network estimated soil surface temperature values for 2 years from the training set of data (1995 and
1996), and for the months of January and July.
G. Mihalakakou / Energy and Buildings 34 (2002) 251259 255
for two randomly selected years from the training set of data
(1995 and 1996) and for the months of January and July,
respectively. As it can be seen from these gures, the
estimated soil temperature values perform well with the
measured ones. For most cases, the soil surface temperatures
are less (18C), while the mean squared error between the
measured and the estimated values was found equal to 0.61
and 0.558C for the month of January and July, respectively.
The correlation coefcients between the estimated and
measured energy consumption values were found equal to
0.94 for the month of January and 0.96 for the month of July.
The temporal variation of the estimated and measured soil
surface temperature values for three, randomly selected
from the training set days of the cold period (68th of
January 1993), and for three randomly selected from the
training set days of the warm period (1214th of July 1994),
is shown in Fig. 4. In this gure, the continual line indicates
the measured soil temperature values while the cross sym-
bols indicate the model's estimations. As shown, there is a
good agreement between the estimated and the measured
data. Quite similar performance was observed for the whole
training set of data.
Furthermore, the neural network's estimated soil surface
temperature values were tested by comparing them with
the actual values of a testing set of data, which consists of
the measured hourly soil surface temperature values for the
year 1998. The temporal variation of the estimated from the
network soil temperature values and of the testing set
measured values for one randomly selected day of January
1998 (18th), and of July 1998 (2nd), are presented in Fig. 5.
Again, the continual line indicates the measured energy
consumption values while the point symbols indicate the
model's estimations. As shown, the neural network esti-
mated values perform well on the testing set of measure-
ments. Quite similar performance was observed for the
whole set of the testing data.
Fig. 6 shows the comparison between the estimated
hourly values and the measured ones of soil surface tem-
perature for the testing set of data. Again, the months of
January and July of 1998 were selected for the presentation
of results. The mean square errors are found equal to 0.72
and 0.608C for January and July, respectively. The correla-
tion coefcients are 0.93 and 0.95 for January and July,
respectively. Similar performance is observed for the whole
set of the testing data. The present results are quite encoura-
ging and the neural network approach is found able to
simulate and estimate the soil surface temperature values
time series with sufcient accuracy.
Furthermore, neural network models are used to investi-
gate and analyze the impact of the three input climatic
Fig. 4. Temporal variation of the estimated with the neural network and of the measured soil surface temperature values for three randomly selected from the
training set days of January 1993 (68th), and for three randomly selected from the training set days of July 1994 (1214th).
256 G. Mihalakakou / Energy and Buildings 34 (2002) 251259
parameters on the uctuation of the soil surface temperature.
For this investigation, neural networks are designed and
trained for the soil temperature estimation using only one
parameter as input each time, for the testing time period
(1998). The models' results are analyzed as follows:
v when the three parameters (air temperature, relative
humidity, and global solar radiation) are used as inputs
to the neural network models, the correlation coefficient
between measured and estimated soil temperature values
was found equal to 0.95;
v when ambient air temperature is the only input parameter,
the correlation coefficient was found equal to 0.68;
v when relative humidity is the only input parameter, the
correlation coefficient was equal to 0.49;
v when global solar radiation is the only input parameter,
the correlation coefficient was equal to 0.56.
The use of one parameter as input to the models, results in
a remarkable reduction of the correlation coefcient which
can be described as follows:
v For the use of ambient air temperature as input parameter
representing the convective energy exchanged between
the air and the ground surface, the correlation coefficient
decrease was approximately 28%.
v For the use of relative humidity as input parameter
representing the latent heat flux, the correlation coeffi-
cient reduction was around 48%.
v For the use of global solar radiation as input parameter
representing the short-wave radiation, the correlation
coefficient reduction was 41%.
The above results showed that the agreement between
estimated and measured values of the soil surface tempera-
ture is better when the ambient air temperature is used as the
unique input to neural models. This can be explained by the
fact that the air temperature is more representative of the
environment physical processes inuencing the soil surface
temperature.
4. Comparison of the analytic and the
data-driven model
The soil surface temperature values estimated by the
two models described in the present paper, the determi-
nistic model and the neural network one, were compared
with the corresponding measured values at the station of
NOA. The comparison was performed for the year 1998,
which was the testing year for the data-driven model. The
Fig. 5. Temporal variation of the estimated with the neural network and of the measured soil surface temperature values for one randomly selected from the
testing set day of January 1998 (18th), and for one randomly selected from the training set day of July 1998 (2nd).
G. Mihalakakou / Energy and Buildings 34 (2002) 251259 257
relative error (R.E.%), was used for the presentation of
results.
(R:E:%) =
R
meas
R
est
R
meas
100
where R
meas
and R
est
are the measured and estimated values
from the models soil surface temperature, respectively.
For the month of July the analytic model estimations are
in close agreement with the measured data with the relative
error uctuating between 6.6 and 4.6%. Quite similar good
performance was also observed from the comparison of
neural network model estimations with the measured values
of soil surface temperature. In this case, the R.E. varied in
the range of 7.8 to 6.7%. The performance of the analytic
model is slightly better as it takes into account a large
number of input parameters. However, the data-driven
model can provide quite similarly good results for the warm
period of the year. Summer period in Athens consists of
clear warm days characterized by the absence of weather
phenomena.
The monthly mean day of January was selected for
representing the cold period of the year. January in Athens
is a month characterized by a great number of cloudy days
and various weather phenomena such as cloud coverage,
rainfall and storms. For the analytic model the R.E. varied
between 7.3 and 8.5%. For the neural network estimations
the R.E. ranged from 14.5 to 12.2%, respectively. For
the cold period of the year the deterministic model gives
better estimations than the data-driven model. This can be
explained by the fact that the analytic model includes a big
number of input parameters and contains several expressions
describing various physical processes inuencing the soil
surface temperature. On the other hand, neural network
models do not use so many input parameters and cannot
simulate with similar success winter days with weather
phenomena as their results depend strongly on the training
data.
5. Conclusions
Soil surface temperature prole in Athens were estimated
in the present study using two models:
v An analytic deterministic model based on the transient
heat conduction differential equation and using as bound-
ary condition the energy balance equation at the ground
surface. The energy balance equation involves the con-
vective energy exchange between air and soil, the solar
radiation absorbed by the ground surface, the latent heat
Fig. 6. Comparison of the measured with the neural network estimated energy consumption values for the year of the testing set of data (1998), and for the
months of January and July.
258 G. Mihalakakou / Energy and Buildings 34 (2002) 251259
flux due to evaporation at the ground surface as well the
long-wave radiation.
v A neural network model based on backpropagation algo-
rithm for the estimation of hourly values of soil surface
temperature. The data-driven model uses as inputs several
climatic parameters such as the ambient air temperature,
the relative humidity and solar radiation. Training and
testing soil temperature values were compared with the
measured ones and it was found that the neural network
can successfully simulate the soil surface temperature
parameter.
Furthermore, the inuence of each one of the input
climatic parameters on the soil surface temperature was
investigated using the neural network models with the
following results:
1. The use of ambient air temperature as a single input
parameter to the neural models produced a decrease of
the correlation coefficient between measured and
estimated values, which arrived at 28%.
2. The use of relative humidity as the only input to the
neural models resulted in a higher reduction of
correlation coefficient, which has risen to 48%.
3. The global solar radiation as a single input parameter
caused a reduction of correlation coefficient equal to
41%.
Furthermore, the estimated values from the two models
and the measured ones were compared and it was found that
the analytic model gives slightly better estimations than
the data-driven model for the warm period of the year. For
the cold period, the analytic model gives obviously better
estimations as it takes into account a large number of input
parameters describing the weather conditions for this period,
which normally is characterized by various weather phe-
nomena such as cloudiness, rainfalls and storms.
References
[1] K. Labs, Ground cooling, in: J. Cook (Ed.), Passive Cooling, MIT
Press, Cambridge, MA, pp. 206250.
[2] A. Tombazis, A. Argiriou, M. Santamouris, Performance evaluation
of passive and hybrid cooling components for a hotel complex,
International Journal of Solar Energy 9 (1990) 112.
[3] G. Mihalakakou, M. Santamouris, D. Asimakopoulos, Modeling the
thermal performance of earth to air heat exchangers, Solar Energy 53
(1994) 301307.
[4] M. Santamouris, G. Mihalakakou, C. Balaras, A. Argiriou, D.
Asimakopoulos, Use of buried pipes for energy conservation in
cooling of agricultural greenhouses, Solar Energy 55 (1995) 111124.
[5] E. Penrod, J.M. Elliot, W.K. Brown, Soil temperature variation at
Lexington, Kentucky Journal of Soil Sciences 90 (1960) 275283.
[6] J.E. Carson, Analysis of soil and air temperatures by Fourier
techniques, Journal of Geophysical Research 68 (1963) 22172232.
[7] T. Kusuda, The effect of ground cover on earth temperature, in:
Proceedings of a 1975 Conference on Alternatives in Energy
Conservation: The Use of Earth-Covered Buildings, Fort Worth,
TX, pp. 912.
[8] B. Givoni, L. Katz, Earth temperatures and underground buildings,
Energy and Buildings 8 (1985) 1525.
[9] G. Mihalakakou, M. Santamouris, D. Asimakopoulos, Modeling the
earth temperature using multiyear measurements, Energy and
Buildings 9 (1992) 19.
[10] P. Bloomfield, Fourier Analysis of Time Series: An Introduction,
Wiley, New York, 1976.
[11] C.P. Jacovides, G. Mihalakakou, M. Santamouris, J.O. Lewis, On the
ground temperature profile for passive cooling applications in
buildings, Solar Energy 57 (1996) 167175.
[12] K. Chakraborty, K. Mehrotra, C.K. Mohan, S. Ranka, Forecasting the
behavior of multivariate time series using neural networks, Neural
Networks 5 (1992) 961970.
[13] G. Mihalakakou, M. Santamouris, D. Asimakopoulos, Modeling the
ambient air temperature time series using neural networks, Journal of
Geophysical Research 103 (19) (1998) 509519.
[14] M. Santamouris, G. Mihalakakou, B. Psiloglou, G. Eftaxias, D.N.
Asimakopoulos, Modeling the global solar radiation on the earth's
surface using atmospheric deterministic and intelligent data-driven
techniques, Journal of Climate 12 (1999) 31053116.
[15] H.S. Carslaw, J.C. Jaeger, Conduction of Heat in Solids, 2nd Edition,
Oxford University Press, Oxford, 1980.
[16] G. Mihalakakou, M. Santamouris, J.O. Lewis, D. Asimakopoulos, On
the application of the energy balance equation to predict ground
temperature profiles, Solar Energy 60 (1997) 181190.
[17] M. Krarti, C. Lopez-Alonso, D.E. Claridge, J.F. Kreider, Analytical
model to predict annual soil surface temperature variation, Journal of
Solar Energy Engineering 117 (1995) 9199.
[18] M. Kreith, S.M. Bohn, Principles of Heat Transfer, West Publishing
Company, 1993.
[19] R.W. Bliss, The derivation of several plate-efficiency factors useful in
the design of flat plate heat collectors, Solar Energy 3 (1959) 5564.
[20] R. Geiger, The Climate Near Ground, Harvard University Press,
Cambridge, MA, 1961.
[21] G. Clark, C. Allen, The estimation of atmospheric radiation for clear
and cloudy skies, in: Proceedings of the 2nd National Passive Solar
Conference, 1978, pp. 675678.
[22] H.L. Penman, Natural evaporation from open water, bare soil and
grass, Proceedings of Royal Society, London A 193 (1948) 120145.
[23] D.T. Pham, X. Liu, Neural Networks for Identification, Prediction,
and Control, Springer, New York, 1995.
[24] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning internal
representations by error propagation, in: D.E. Rumelhart (Ed.), Parallel
DistributedProcessing, MITPress, Cambridge, MA, 1986, pp. 318362.
G. Mihalakakou / Energy and Buildings 34 (2002) 251259 259

Das könnte Ihnen auch gefallen