See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/248701454

2010 - Estimation of static formation temperatures in geothermal

Data · July 2013

CITATIONS READS
0 233

4 authors, including:

Alfonso Garcia-Gutierrez Gilberto Espinosa-Paredes

Instituto de Investigaciones Electricas Metropolitan Autonomous University-Iztapalapa
129 PUBLICATIONS   689 CITATIONS    209 PUBLICATIONS   1,215 CITATIONS   

SEE PROFILE SEE PROFILE

E. Santoyo
Universidad Nacional Autónoma de México
119 PUBLICATIONS   1,461 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Neutron View project

LCA studies for renewable energy technologies View project

All content following this page was uploaded by E. Santoyo on 20 May 2014.

The user has requested enhancement of the downloaded file.

 Computers & Geosciences 36 (2010) 1191–1199

Contents lists available at ScienceDirect

Computers & Geosciences

journal homepage: www.elsevier.com/locate/cageo

Estimation of static formation temperatures in geothermal wells by using an

artificial neural network approach
A. Bassam a, E. Santoyo a,n, J. Andaverde b, J.A. Hernández b, O.M. Espinoza-Ojeda a
a
Centro de Investigación en Energı́a, Universidad Nacional Autónoma de México, Privada Xochicalco s/n, Centro, 62580 Temixco, Morelos, México
b
Centro de Investigación en Ingenierı́a y Ciencias Aplicadas, Universidad Autónoma del Estado de Morelos, Av. Universidad 1001, Col. Chamilpa, 62209 Cuernavaca, Morelos, México

a r t i c l e in fo abstract

Article history: An artificial neural network (ANN) approach was used to develop a new predictive model for the
Received 13 August 2009 calculation of static formation temperature (SFT) in geothermal wells. A three-layer ANN architecture
Received in revised form was successfully trained using a geothermal borehole database, which contains ‘‘statistically normal-
6 January 2010
ised’’ SFT estimates. These estimates were inferred from seven analytical methods commonly used in
Accepted 7 January 2010
geothermal industry. Bottom-hole temperature (BHT) measurements and shut-in times were used as
main input variables for the ANN training. Transient temperature gradients were used as secondary
Keywords: variables. The Levenberg–Marquardt (LM) learning algorithm, the hyperbolic tangent sigmoid transfer
Horner method function and the linear transfer function were used for the ANN optimisation. The best training data set
Geothermal energy
was obtained with an ANN architecture composed by five neurons in the hidden layer, which made
Levenberg–Marquardt algorithm
possible to predict the SFT with a satisfactory efficiency (R2 4 0.95). A suitable accuracy of the ANN
Borehole drilling
Bottom-hole temperature model was achieved with a percentage error less than 7 5%. The SFTs predicted by the ANN model were
Artificial intelligence statistically analyzed and compared with ‘‘true’’ SFTs measured in synthetic experiments and actual
Shut-in time BHT logs collected in geothermal boreholes during long shut-in times. These data sets were processed
both to validate the new ANN model and to avoid bias. The SFT estimates inferred from the ANN
validation process were in good agreement (R2 40.95) with the ‘‘true’’ SFT data reported for synthetic
and field experiments. The results suggest that the new ANN model could be used as a practical tool for
the reliable prediction of SFT in geothermal wells using BHT and shut-in time as input data only.
& 2010 Elsevier Ltd. All rights reserved.

1. Introduction The determination of the static formation temperatures (SFTs)

from BHT measurements constitutes a crucial task for the
The exploitation of geothermal resources for producing evaluation of geothermal systems (Espinosa et al., 2001; Kutasov
electricity requires the drilling of deep boreholes in the most and Eppelbaum, 2003). The estimation of SFT from BHT data at
suitable thermal regions of geothermal fields (Saito et al., 1998; early drilling times offers the chance to know the virgin formation
Davis and Michaelides, 2009). The borehole drilling is a complex temperature several months before it can be measured with
process in which a constant thermal anomaly (added to a accuracy. This approach is valuable in terms of planning,
circulating drilling mud) affects the original rock-formation exploration, developing, and exploitation programmes. The
temperature surrounding the borehole (Fomin et al., 2003, knowledge of the SFT is required for the: (1) determination of
2005). Once the borehole drilling is completed, the resulting geothermal gradients for exploration mapping; (2) determination
thermal recovery is evaluated by analyzing build-up bottom-hole of heat flows of promising geothermal zones; (3) interpretation of
temperature (BHT) and shut-in time measurements (Santoyo borehole logging; (4) optimal design of borehole drilling and
et al., 2000). BHT data are measured at different shut-in times completion programs; (5) location of permeable or lost circulation
during the borehole drilling operations (Espinosa-Paredes and zones; (6) evaluation of in-situ thermophysical formation proper-
Espinosa-Martinez, 2009). BHTs are usually costly due to the use ties; and (7) for the estimation of the heat reserves in geothermal
of sophisticated logging equipment, but mostly, because the systems.
borehole drilling progress needs to be stopped (Wisian et al., The estimation of SFT in geothermal boreholes is usually
1998; Fomin et al., 2005). performed by using simplified analytical methods based on
complex heat transfer models concentrated at the bottom-hole
conditions, where BHTs are actually measured. The BHT measure-
n
Corresponding author. ments tend to reflect the thermal anomalies caused by the drilling
E-mail address: esg@cie.unam.mx (E. Santoyo). mud circulation to the rock-formation.

0098-3004/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2010.01.006
1192 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199

The most common analytical methods include: (i) the radial consist of BHT and shut-in time measurements, and transient
source with a conductive heat flow or Brennand method (BM: temperature gradients, together with SFT estimates inferred from
Brennand, 1984); (ii) the cylindrical heat source with a con- the regression analysis of curvilinear relationships between BHT
ductive-convective heat flow method (CSM) proposed by Hasan and shut-in time data using simplified analytical methods. SFT
and Kabir (1994); (iii) the line-source or the well-known Horner- estimates were calculated with seven analytical methods com-
plot method (HM: Dowdle and Cobb, 1975); (iv) the generalized monly used in geothermal industry.
Horner or the Kutasov–Eppelbaum method (KEM: Kutasov and The normalisation of SFT estimates was carried out using
Eppelbaum, 2005); (v) the cylindrical source with a conductive statistical tests for discordant outlier detection and rejection
heat flow or Leblanc method (LM: Leblanc et al., 1981); (vi) the (Verma, 2005). Mean SFT estimates (normalised) and standard
cylindrical source with a conductive heat flow or Manetti method deviations were computed using an improved outlier detection/
(MM: Manetti, 1973); (vii) the spherical and radial heat flow rejection statistical method (Verma et al., 2009). This statistical
method (SRM) proposed by Ascencio et al. (1994); among others. approach coupled to the ANN enables the resulting errors to be
Most of these analytical methods require at least three or more effectively detected and rejected before selecting reliable SFT
BHT measurements carried out at the same borehole depth but at estimates. The new predictive ANN model was evaluated by using
different shut-in times. The SFT is determined through the four BHT data sets (two synthetic and two actual borehole data
mathematical solution of each analytical method, using BHT and sets) measured in experimental and long logging works where the
shut-in time measurements as input data, and the linear or SFT or the true formation temperature (TFT) was already known.
nonlinear regressions between BHT data and the time functions of These validation tests were considered as an accuracy evaluation
each method (Verma et al., 2006a, b). of the ANN model.
Although considerable progress has been achieved in this area
over recent years, large discrepancies among the SFT results
2.1. Geothermal borehole databases for learning and validation tests
provided by different methods have been reported (Andaverde
et al., 2005). Therefore, the development of new reliable methods
A geothermal borehole database containing BHT and shut-in
to estimate SFT is still a challenging area. The geothermal industry
time (Dt) measurements logged from drilling operations of 10
is interested in estimating SFT in systems where heat convection
boreholes was created (Table 1). These thermal recovery data sets
dominates. Heat convection is attributed to the presence of
were collected from the reports of boreholes drilled in various
permeable zones in reservoirs and drilling mud losses into the
world geothermal fields: (1) Los Humeros, Mexico [MXCO1,
formation. The determination of SFT under such conditions
number of BHT measurements (mBHT)¼6; shut-in times (Dt) up
requires the use of complex analytical methods or wellbore
to 36 h after cessation of drilling mud circulation (Andaverde
simulators. These tools need thermophysical and transport
et al., 2005)]; (2) Los Humeros, Mexico [MXCO2, mBHT ¼7;
properties of drilling, cementing, and formation materials, which
Dtr42 h (Andaverde et al., 2005)]; (3) Oklahoma, USA [SGIL,
are rarely available in the literature. In addition, the accurate
mBHT ¼13; Dt r12 h (Schoeppel and Gilarranz, 1966)]; (4)
knowledge of the mud circulation time, which is generally
Mississippi wellbore, USA [USAM, mBHT ¼14; Dt r200 h
unknown or difficult to determine under drilling conditions, is
(Kutasov, 1999)]; (5) Leyte, Philippines [PHIL, mBHT ¼14;
also needed by most analytical methods (Hermanrud et al., 1990).
Dtr15.58 h (Andaverde et al., 2005)]; (6) Kyushu, Japan [JAPN,
Considering this complex scenario, the geothermal industry
mBHT ¼10; Dt r72.5 h (Hyodo and Takasugi, 1995)]; (7) Chipilapa,
requires new practical tools that use the commonly available data
El Salvador [CH-A1, mBHT ¼5; Dt r95.27 h (González-Partida
(i.e., BHT and shut-in time) to estimate the SFT with an acceptable
et al., 1997)]; (8) Chipilapa, El Salvador [CH-A2, mBHT ¼6;
accuracy. In this regard, a new application of artificial neural
Dtr190.5 h (Iglesias et al., 1995)]; (9) Larderello, Italy [ITAL,
networks (ANN) is here proposed for developing a new reliable
mBHT ¼40; Dt r27 h (Da-Xin, 1986)]; and (10) Norton Sound field,
method for the calculation of SFTs in geothermal wells.
Alaska [COST, mBHT ¼ 6; Dt r75.5 h (Cao et al., 1988a)].
ANN have been suggested as powerful computational tools in
Four thermal recovery data sets were additionally compiled for
many sciences for modelling and solving complex real-world
the validation or accuracy tests:
problems, and mainly for forecasting applications (Zhang et al.,
(i) two synthetic data sets collected from experimental works
1998). The use of ANN is increasingly in Earth Sciences (e.g.,
reported in the literature (SHBE, mBHT ¼8: Shen and Beck, 1986;
Goutorbe et al., 2006; Hsieh et al., 2009; Leite and Filho, 2009a, b;
and CLAH, mBHT ¼15: Cao et al., 1988b); and (ii) two data sets
Morton, 2009). In recent years, the application of ANN to
logged in some geothermal boreholes which were collected
geothermal and petroleum engineering problems has also been
during long logging works (CH-A4 borehole, mBHT ¼5: Iglesias
the subject of study (e.g., Farshad et al., 2000; Bayram, 2001; Can,
et al., 1995; and KJ-21 borehole, mBHT ¼6: Steingrimsson
2002; Yilmaz et al., 2002; Spichak and Goidina, 2005, Spichak,
and Gudmundsson, 2006). These data sets were considered
2006; Dı́az-González et al., 2008; Serpen et al., 2009).
for the evaluation of accuracy because the TFTs were already
In this work, a three-layer ANN approach was successfully
known (SHBE¼80.0 1C; CLAH¼120.0 1C; CH-A4¼169 1C; and
trained for obtaining a new predictive model to estimate SFT in
KJ-21¼240 1C, respectively).
geothermal wells. The new ANN model requires BHT and shut-in
time measurements (as main input data), normalised SFT
estimates (inferred from seven analytical methods), and transient 2.2. Initial estimation of SFTs using analytical methods
temperature gradients.
Seven analytical methods based on complex heat transfer
models were used for the initial calculation of SFTs. The seven
2. Methodology analytical methods were: (i) the Horner-plot method, HM; (ii) the
Brennand method, BM; (iii) the Kutasov–Eppelbaum method,
The ANN model was trained with a geothermal borehole KEM; (iv) the Leblanc method (LM); (v) the Manetti method
database containing SFT estimates (statistically normalised). The (MM); (vi) the conductive spherical-radial method (SRM);
geothermal database also contains build-up thermal recovery logs and (vii) the conductive-convective cylindrical source method
recorded during drilling operations carried out in some geother- (CSM). The simplified solutions of these analytical methods are
mal boreholes of the world. Build-up thermal recovery logs summarised in Table 2. SFTs were estimated using conventional
 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199 1193

Table 1
Thermal recovery data sets used for ANN training and validation.

Data set Borehole data for training the new ANN model Synthetic data for the validation

t (h) BHT (1C) Data set t (h) BHT (1C) Data set t (h) BHT (1C) Data set t (h) BHT (1C)

MXCO1 6 186.3 PHIL 2.58 93 ITAL 5.7 84.5 SHBE 2.5 56.6
tc ¼ 2.5 h 12 206.9 tc ¼15 h 3.58 88 tc ¼ 5 h 6.0 93.1 tc ¼ 5 h 5 61.3
m ¼6 18 219.0 m¼14 4.58 99 m¼ 40 6.5 95.0 m ¼8 7.5 64.3
24 231.9 5.58 108 7.0 96.2 10 66.6
30 239.8 6.58 112 7.5 97.5 15 69.6
36 247.7 7.58 117 8.0 98.4 20 71.7
8.58 120 8.5 99.3 30 74.1
MXCO2 6 178.6 9.58 126 9.0 100.3 40 75.5
tc ¼ 2.5 h 12 198.9 10.58 133 9.5 101.1
m ¼7 18 211.4 11.58 133 10.0 102.0
24 225.8 12.58 134 10.5 102.7
30 235.1 13.58 137 11.0 103.4 CLAH 2 91.7
36 240.4 14.58 141 11.5 104.2 tc ¼ 5 h 4 98.1
42 247.1 15.58 146 12.0 104.7 m ¼15 6 102.4
12.5 105.4 8 105.5
SGIL 0 62.78 JAPN 5.5 92.0 13.0 106.1 10 107.9
tc ¼ 3 h 1 81.94 tc ¼5 h 6.5 98.5 13.5 106.7 12 109.8
m ¼13 2 86.56 m¼10 7.5 103.0 14.0 107.6 14 111.3
3 88.84 8.5 107.4 14.5 108.3 16 112.5
4 90.76 9.5 110.4 15.0 108.9 18 113.6
5 92.29 12.5 119.6 15.5 109.6 20 114.4
6 93.44 15.5 126.6 16.0 110.2 22 115.2
7 94.21 18.5 132.8 16.5 110.8 24 115.8
8 94.59 24.5 142.4 17.0 111.3 30 117.1
9 94.98 72.5 170.9 17.5 111.9 40 118.4
10 95.36 18.0 112.4 50 119.1
11 95.74 CH-A1 6 90 18.5 112.9
12 96.13 tc ¼5 h 12 102 19.0 113.3 CH-A4 6 85
m¼ 5 26.57 116 19.5 113.8 tc ¼ 5 h 12 94
47.1 130 20.0 114.2 m ¼5 26.57 122
USAM 1 138.75 95.27 145 20.5 114.7 47.1 139
tc ¼ 5 h 2 139.55 21.0 115.1 95.27 152
m ¼14 5 140.89 CH-A2 6.0 231 21.5 115.4
10 142.03 tc ¼5 h 12.0 231 22.0 115.7 KJ-21 2 50
15 142.72 m¼6 26.6 235 22.5 116.1 tc ¼ 2.5 h 30 125
20 143.21 47.1 237 23.0 116.5 m ¼6 52 158
25 143.60 95.3 239 24.0 117.1 96 184
30 143.92 190.5 240 25.0 117.7 242 218
40 144.42 26.0 117.9 556 225
50 144.81 COST 8.5 45.00 27.0 118.7
70 145.40 tc ¼4 h 10.0 49.44
100 146.04 m¼6 16.5 55.56
150 146.76 34.5 56.11
200 147.27 38.0 57.22
75.5 56.11

tc—drilling mud circulation time; additional data reported for each data set: SGIL (R ¼0.10028, k¼2.25419 W/m 1C, c¼ 878.64 J/kg 1C, and r ¼ 2306.6 kg/m3); COST
(k¼ 4.97 W/m 1C); SHBE (r ¼0.108 m, TFT¼ 80 1C); CLAH (r ¼0.108 m, TFT ¼120 1C); CH-A4 (TFT¼ 169 1C); KJ-21(TFT¼ 240 1C). m—number of data.

Table 2 linear (ordinary and weighted) and quadratic (QR) regressions

Listing of major numerical solutions of analytical methods (HM, BM, KEM, LM, between the BHT and the time function data of each method (see
MM, SRM, and CSM) used for determining SFT in geothermal boreholes. Andaverde et al., 2005). Ordinary linear regressions (OLR) neglect
the errors in the x (time functions) and y (BHT) variables for
Analytical method Numerical solution
the calculation of the regression parameters: intercept (a) and
 
BM BHTðtÞ ¼ TBM bBM 1 slope (b), whereas the weighted linear regressions (WLR) consider
Dt þ ptc
CSM BHT ¼ TCSM + (bCSM)[F(tcD + DtD)  F(tD)]
the errors of x and y (sx and sy) for the estimation of the regression
HM BHTðtÞ ¼ THM þ ðbHM Þlnfðtc þ DtÞ=Dtg parameters (Verma, 2005).
KEM BHTðtÞ ¼ TKEM þ bKEM lnðXÞ According to the numerical procedure used by each analytical
h  i
LM BHTðtÞ ¼ TLM bLM 1exp  4awDt
r2
method, SFTs were inferred from the intercept values (a) of the
  OLR, WLR and QR (e.g., Santoyo et al., 2000; Andaverde et al., 2005).
MM BHTðtÞ  TMM þ bMM ln DttDt
c
 
SRM BHTðtÞ ¼ TSRM þ ðbSRM Þ p1ffiffiffiffiDt
2.3. General scheme for the SFT estimation
BHT: Bottom-hole temperature; Dt: shut-in time; SFTs were inferred from the
intercept value of each analytical method: TBM, TCSM, THM, TKEM, TLM, TMM, and TSRM.
For more details of primary and secondary variables, the readers are referred to A schematic flow diagram showing the training procedure
the original sources of the methods (indicated in the introduction section). used by the ANN model is shown in Fig. 1. Five fundamental
1194 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199

(i)
Input data sets of geothermal
wells (n = 10)

Shut-in
BHT (Temp)
time (t)

(ii)
Selection of analytical methods
and calculation of their time
functions: BM, CSM, HM, KEM,
LM, MM and SRM

(iii)
Estimation of SFT using three
regression models OLR, WLR
and QR

(v) (iv)
Calculation- Estimation of mean and standard
Transient temperature deviation values of SFT using
gradients statistical discordant tests to define
a normal distribution

Mean and standard deviation

values: xSFT ± sSFT

t,T ANN
dT/dt Training model Derivation of
new equation

ANN
Evaluation
Results

Validation

Fig. 1. Schematic flow diagram of ANN model used for determining SFT in geothermal boreholes.

sections were included in the training process: (i) the input data and CSM) were calculated by means of their corresponding
reading (j¼ 10, containing 120 pairs of BHT and shut-in time equations (Table 2). Three SFT estimates were individually inferred
data); (ii) the selection of analytical methods and calculation of from the regression analysis (OLR, WLR, and QR) using each
their time functions per data set; (iii) the determination of the SFT analytical method and referred to each data set. A total number of
by seven analytical methods using three different regression 21 SFT estimates per data set were finally computed.
models: OLR, WLR, and QR per data set; (iv) the application of For training and correlating the borehole thermal recovery
statistical discordant tests to calculate the mean values of the SFT data sets with reliable SFT values, mean values of the 21 SFT
estimates and their standard deviations per data set; and (v) the estimates were computed per data set through a statistical
calculation of temperature gradients using the transient thermal normalisation using the software DODESYS (Verma et al., 2009).
recovery data of the boreholes. Nine discordant tests (N1, N2, N4, N7, N8, N9, N10, N14, and N15
SFT estimates were individually computed by using the 10 data proposed by several authors: see Verma, 2005) were applied to
sets (MXCO1, MXCO2, SGIL, USAM, PHIL, JAPN, CH-A1, CH-A2, ITAL, the 21 SFT estimates of each data set; firstly, to identify and to
and COST), which contain BHT and shut-in time data. The time reject discordant outliers, and secondly, to calculate the mean SFT
functions of each analytical method (HM, BM, KEM, LM, MM, SRM, (xSFT ) and its associated standard deviation (sSFT).
 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199 1195

This method of statistical normalisation enabled the large where BHT denotes the bottom-hole temperatures, t the shut-in
resulting errors derived from the analytical methods to be times, and the subscripts i provides the transient values of these
efficiently detected and rejected before selecting the best SFT variables.
estimates. Table 3 summarises the results obtained after using the The transient temperature gradient (dT/dt) was an important
statistical normalisation procedure. parameter for training the ANN. When (dT/dt) tend to zero at
The mean SFT estimates and discordant outliers detected after infinite time(t-N), it means that the geothermal formation
applying the statistical tests are plotted in Fig. 2. Thirty discordant temperature approaches to thermal equilibrium, and therefore:
outliers were detected and rejected. These outliers corresponded
dT
to the SFT estimates inferred from the SRM which systematically limt-1 ¼ 0 ‘ ðBHT  SFT Þ ð2Þ
dt
overestimate the SFT (Andaverde et al., 2005). The mean values of
the SFT estimates (xSFT ) and their standard deviations (sSFT) per The SFTANN was defined as the unique output parameter of the
data set were used for the ANN training. Three SFT values ANN model.
were finally employed for transporting the uncertainties (the
mean value: xSFT ; the mean value plus the standard deviation:
xSFT þsSFT ; and the mean value minus the standard deviation: 2.4. ANN architecture
xSFT sSFT ) (see Table 3).
Transient temperature gradients were computed using the The neurons are grouped into distinct layers and intercon-
following equation: nected according to a given computational architecture. The
network function is largely determined by the connections
dT ðBHTi þ 1 BHTi Þ between neurons. Each connection among two neurons has a
¼ ð1Þ
dt ti þ 1 ti weighting coefficient. The standard network structure for an
approximation function is the multiple-layer perception (or the
well-known feed forward network). The feed forward network
Table 3
Mean SFT estimates inferred from 21 SFT predictions of seven analytical methods
often has one or more hidden layers of sigmoidal neurons
(using three different regression models OLR, WLR, and QR), and after applying a followed by an output layer of linear neurons. Multiple-layers of
statistical normalisation process. neurons with nonlinear transfer functions allow the network to
learn nonlinear and linear relationships between input and output
Data set SFT (analytical methods) SFT (predicted—ANN)
vectors. The linear output layer enables the network to generate
xSFT 7 sSFT xANN 7 sANN
values outside the –1 to + 1 range (e.g., Limin, 1995). For the
MXCO1 263 7 15 256 7 20 network, the appropriate notation is used in two-layer networks
MXCO2 262 7 18 253 7 15 (e.g. Demuth and Beale, 1998). A simplified sketch of the network
USAM 146 7 1 139 7 5
structure is schematically represented in Fig. 3. The number of
ITAL 132 7 10 133 7 8
PHIL 197 7 28 185 7 25 neurons in the input and output layers is, respectively, given by
JAPN 176 7 18 180 711 the number of input and output variables of the process under
CH-A1 147 7 5 151 7 7 investigation. In this work, the input layer can consist of a set of
CH-A2 241 7 18 243 7 15 different variables, Ink (i.e., BHT, shut-in time, and dT/dt), while
COST 577 3 567 2
the output layer can contain the output or target variables of the
SGIL 100 73 102 79
model, Outl (i.e., SFT), see Fig. 3.

400
Mean
Mean ± sd
350
Outliers
Extremes
300
SFT (analytical methods) °C

250

200

150

100

50

0
MXCO1 MXCO2 USAM ITAL PHIL JAPN CH-A1 CH-A2 COST SGIL
DATA SET

Fig. 2. Mean SFT estimates obtained from normal statistical samples after applying nine statistical discordant tests. Mean SFT estimates per data set were: MXCO1:
263 7 15 1C; MXCO2: 262 718 1C; USAM: 146 7 1 1C; ITAL: 132 7 10 1C; PHIL: 1977 28 1C; JAPN: 176 7 18 1C; CH-A1: 1477 5 1C; CH-A2: 241 7 18 1C; COST: 57 73 1C;
and SGIL: 1007 3 1C.
1196 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199

Newton’s method (Eq. (6)), which is suggested to approach a

second order training speed without having to compute the
Hessian matrix (Hagan and Menhaj, 1994)
 1
xk þ 1 ¼ xk  J T J þ mI J T e ð6Þ

where J is the Jacobian matrix that contains first derivatives of the

network errors with respect to the weights and biases; e is the
vector of network errors; JTJ is an approximation of the Hessian
matrix; and JTe is the gradient. When the scalar m is zero, Eq. (6)
becomes the classical Newton’s method using the approximate
Hessian matrix. On the other hand, when m is large, Eq. (6)
becomes a more efficient descent gradient with a small step size.
The Jacobian matrix can be computed through a standard back-
Fig. 3. General ANN architecture (k¼ input number; In¼ Input; Out ¼ Output; propagation technique (see Hagan and Menhaj, 1994).
Continuous lines ¼weights and bias). Despite the fact that computations involved in iterations
were more complex than the steepest descent case, the conver-
The optimal number of neurons in the hidden layer (S) is gence was faster (typically by a factor of 100). The root mean
difficult to specify, and it generally depends on the problem square error (RMSE) was calculated using the experimental and
complexity. This number (S) is iteratively determined. Each simulated ANN data (Eq. (7)). This calculation was used as an
neuron in the hidden layer has a bias, b1(s), which is added to optimisation criterion for the model adequacy (Fig. 4).
the sum of the product (weights multiplied by its corresponding The experimental database (see Table 1) were split into
input variables). In general, the optimal neuron (ns), is given by learning (60% of experimental data set; n ¼180) and testing
the following equation: (40% of experimental data set; n ¼120) databases to obtain a
ns ¼ Wifs,1g In1 þWifs,2g In2 þ    þ Wifs,kg Ink þ b1 ðsÞ ð3Þ statistically representative sample of the problem
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n
The coefficients associated with the hidden layer are grouped u1 X
into a matrix Wi (given by weights) and the vector b1(s) of biases. RMSE ¼ t ðExpðiÞSimðiÞÞ2 ð7Þ
ni¼1
The output layer computes the weighted sum of the signals
provided by the hidden layer, and the associated coefficients are
grouped into a matrix Wo and the vector b2(l). Using the matrix
3.1. ANN training
notation, the network output can be expressed by the following
general equation:
The new neural network model trained for this work (Fig. 5)
Out ¼ g ðWo f ðWi In þ b1 Þ þb2 Þ ð4Þ finally involved five neurons (ns ¼ 5) in the hidden layer with
Hidden layer neurons may use any differentiable transfer 20 weights (15 elements of a matrix Wi and 5 elements of the
function to generate their output. In this work, a hyperbolic matrix Wo), and 6 biases (b1 ¼5 and b2 ¼1) to calculate a SFT value
2 0 13
tangent sigmoidal transfer function (f¼tansig) and a linear
transfer function (g¼purelin) were used. As a result, Eq. (4) can XS 6
6
B
B
C7
C7
6Wo ð1,jÞB 2
be re-written using these transfer functions to provide the SFT ¼ 6 B !! 1C 7
C7 þ b2
j¼14 @ P
K A5
following general equation: 1þ exp 2 ðWi ðj,kÞInðkÞÞ þ b1 ðjÞ
2 0 13 k¼1

ð8Þ
XS 6
6
B
B
C7
C7
6Wo ðl,sÞB 2
OutðlÞ ¼ !! 1C 7
C7 þ b2ðlÞ
6 B
P
K where S and K were already defined and l¼1 (see Table 4).
s¼14 @ A5
1 þ exp 2 ðWi ðs,kÞInðkÞÞ þ b1 ðsÞ During the learning process, the algorithm of the ANN model
k¼1
was evaluated by examining the RMSE value per iteration,
ð5Þ
together with a different number of neurons in the hidden layer
where K, S, and l are the number of input variables, the number of (ranging from one to five neurons). The obtained results (data not
neurons in the hidden layer, and the number of the output shown) proved that the typical learning error decreased when the
variables in the ANN architecture, respectively. number of neurons in the hidden layer increased. This observation
was evident because of the number of adjusted parameters
increased. Nevertheless, one of the problems that could occur
3. Neural network learning and testing during feed-forward neural network training is the well-known
‘‘over-fitting’’ (e.g. Hernández-Pérez et al., 2004).
A learning (or training) algorithm is defined as a process that The comparison of the RMSE calculated for the learning and
consists of adjusting the coefficients (weights and biases) of a testing databases was a good criterion to optimize the number of
network, to minimize an error function (usually a quadratic one) iterations, and to avoid the over-fitting problem. In this neural
between the network outputs (for a given set of inputs) and the network, the RMSE showed that for six neurons in the hidden
correct outputs (already known). If smooth non-linearities are layer, the learning database value was too small with respect to
used, the error function gradient can be computed by the classical the testing database.
back propagation model (Rumelhart et al., 1986). According to RMSE results, the optimal number of neurons in
Previous learning algorithms have used this gradient directly the hidden layer was fixed as five neurons (Fig. 5). The predicted
in a steepest descent optimisation, but recent results have shown SFT values computed with the ANN model together with their
that second order methods are far more efficient. In this work, the uncertainties are also included in Table 3. A good agreement
Levenberg–Marquardt (LM) optimisation algorithm (described in between the SFT estimates (inferred from analytical methods:
the Matlab Neural Network Toolbox) was used (Demuth and x-axis) and the simulated data, SFTANN (obtained from the ANN
Beale, 1998). This algorithm is basically an approximation of model: y-axis) was found (Fig. 6). It can be observed that the
 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199 1197

Geothermal Analytical Method

well and SFTExp
Statistical Treatment
+

Temperature RMSE Optimization method

Time (Levenberg-Marquardt)
dT/dt
-

Neural Network
SFTSim
(Weights and Bias)

Fig. 4. Numerical algorithm of ANN used for learning and optimizing processes.

Fig. 5. ANN architecture used for estimating SFT in geothermal boreholes.

simulated SFTANN data exhibit a good linear relationship with 3.2. ANN testing
respect to those estimates inferred from analytical methods
(y¼0.94376x + 8.5631; R2 ¼0.95). This linear relationship was The new ANN model (Eq. (8)) was validated by using the
also confirmed by analyzing the small residuals computed testing data reported for synthetic experiments (SHBE: TFT¼80 1C;
between the measured and predicted SFT values. and CLAH: true TFT¼120 1C); and for actual geothermal borehole
The optimum weighting coefficients (Wi, Wo, b1, and b2) measurements (CH-A4: TFT¼169 1C; and KJ-21: true TFT¼240 1C)
obtained for the ANN model using five neurons in the hidden (see Table 1). The results from these four accuracy tests showed a
layer are reported in Table 4. These coefficients were used to good agreement (less than a 75%) between the measured SFT
validate the ANN model with testing data. (also called TFT) and predicted SFT data (Fig. 7).
1198 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199

Table 4
Optimised weighting and bias parameters of new ANN model.

ANN parameters With l¼ 1, S ¼5 and K ¼3

Wi 3.9757 1.4220  2.9996

5.0716 13.9674  2.3583
 6.9071  2.7251  34.0318
 6.3881 2.3519  6.0095
 5.2657 3.6209  4.5029

Wo  1.4449 0.1627  0.3257 0.4320  0.5981

b1  5.6937
 10.0967
1.6998
 0.5554
 2.6091

b2  1.4001
Fig. 7. Evaluation of ANN accuracy by using a statistical comparison between ANN
simulated SFT values and true SFT data or TFT (reported in synthetic and actual
borehole data sets).

Other additional advantages of the ANN model included: (i) to

use the ANN tool without any assumptions about the nature of
underlying mechanisms; (ii) to have the ability to take into
account non-linearities and iterations between input and output
variables; (iii) to optimise the computing time because these
models rely on simple mathematical operations (allowing faster
calculations); and (iv) to obtain reliable and faster predictions of
SFT in comparison with those inferred from current analytical
methods.

4. Conclusions

A new ANN model was successfully developed to predict the

SFT in geothermal boreholes. It only requires BHT and shut-in
time measurements as main input data. This feature makes it an
advantageous tool over current available analytical methods that
strongly depend on other complex input variables, such as the
drilling mud circulation time. The ANN model was effectively
trained with an experimental database, and validated with an
Fig. 6. Statistical comparison between SFT estimates (inferred from analytical unbiased testing database. The validity of the SFT computed by
methods) and ANN simulated SFT values. the ANN model was confirmed through a statistical comparison
between the measured and simulated values, showing a good
agreement with errors less than 75%.
For the synthetic data sets, the ANN model predicted SFT The new ANN model can be used as a practical tool for the
values ranging from 74.5 to 83.4 1C, with a mean SFTANN value of reliable determination of SFT in geothermal boreholes, especially
79.1 1C for the SHBE data set; and SFT values in the interval when the BHT data are in the training interval (45 to 263 1C).
117.8–125.6 1C, with a mean SFTANN value of 122.1 1C for the CLAH The ANN application interval could be extended by using higher
data set. Whereas for the borehole data sets: CH-A4 and KJ-21, the BHT measurements. The predicting accuracy of the ANN could
ANN model predicted SFT values in the interval 167.6–173.5 1C, also be improved by using training patterns with a large number
with a mean SFTANN value of 171.2 1C; and SFT values ranging of BHT measurements, and their uncertainties.
from 239.1 to 247.1 1C, with a mean SFTANN value of 243.3 1C, The idea of applying ANN for the estimation of SFT is to
respectively. propose a new practical and faster tool for the geothermal
From these results, it can be reliably demonstrated that the industry. In fact, the ANN model here developed could be coupled
new ANN model (given by the Eq. (8)) provides acceptable with smart BHT sensors for on-line determination of SFT in
estimates of the SFT with typical errors commonly reported in geothermal boreholes.
artificial intelligence engineering applications (e.g. Leite and Filho, Finally, the ANN model developed may enable the implemen-
2009a). tation of smart sensors for on-line determination (SFT) in
The small residuals calculated with the testing database, as geothermal boreholes.
well as the good regression coefficient obtained (R2 ¼ 0.95),
enabled to confirm the satisfactory agreement between the
measured and predicted SFT data (see Fig. 7). The ANN model Acknowledgements
developed was therefore able to predict the SFT throughout
the experimental data domain used for the training process The authors wish to thank to the Engineering Ph.D. Programme
(45–263 1C). of UNAM and to CONACyT for the financial support provided.
 A. Bassam et al. / Computers & Geosciences 36 (2010) 1191–1199 1199

Thanks are also due to Dr. A. Santoyo-Castelazo (University Hyodo, M., Takasugi, S., 1995. Evaluation of the curve-fitting method and the
of Liverpool) for her valuable comments. We are grateful to Horner plot for estimation of the true formation temperature using
temperature recovery logging data. In: Proceedings of the 20th Workshop
the anonymous reviewer, Dr. Viacheslav V. Spichak, and on Geothermal Reservoir Engineering. Stanford University, Stanford, USA,
Prof. Dr. E.C. Grunsky (Editor in Chief) for their helpful comments pp. 94–100.
on an earlier version of this paper. Iglesias, E.R., Campos-Romero, A., Torres, R.J., 1995. A reservoir engineering
assessment of the Chipilapa, El Salvador, geothermal field. In: Proceedings of
the World Geothermal Congress, Florence, Italy, pp. 1531–1536.
References Kutasov, I.M., 1999. Applied Geothermics for Petroleum Engineers. Elsevier,
Amsterdam 360 pp.
Kutasov, I.M., Eppelbaum, L.V., 2003. Prediction of formation temperatures in
Andaverde, J., Verma, S.P., Santoyo, E., 2005. Uncertainty estimates of static
permafrost regions from temperature logs in deep wells-field cases. Perma-
formation temperature in borehole and evaluation of regression models.
frost and Periglacial Process 14, 247–258.
Geophysical Journal International 160, 1112–1122.
Kutasov, I.M., Eppelbaum, L.V., 2005. Determination of formation temperature
Ascencio, F., Garcı́a, A., Rivera, J., Arellano, V., 1994. Estimation of undisturbed
from bottom-hole temperature logs—a generalized Horner method. Journal of
formation temperatures under spherical radial heat flow conditions. Geother-
Geophysical Engineering 2, 90–96.
mics 23, 317–326.
Leblanc, Y., Pascoe, L.J., Jones, F.W., 1981. The temperature stabilization of a
Bayram, A.F., 2001. Application of an artificial neural network model to a Na/K
borehole. Geophysics 46, 1301–1303.
geothermometer. Journal of Volcanology and Geothermal Research 112, 75–81.
Leite, E.P., Filho, C.R.de S., 2009a. Probabilistic neural networks applied to mineral
Brennand, A.W., 1984. A new method for the analysis of static formation
potential mapping for platinum group elements in the Serra Leste region,
temperature test. In: Proceedings of the 6th New Zealand Geothermal
Carajás mineral province, Brazil. Computers & Geosciences 35, 675–687.
Workshop, Auckland, New Zealand, pp. 45–47.
Leite, E.P., Filho, C.R.de S., 2009b. TEXTNN—a MATLAB program for textural
Can, I., 2002. A new improved Na/K geothermometer by artificial neural networks.
classification using neural networks. Computers & Geosciences 35, 2084–2094.
Geothermics 31, 751–760.
Limin, F., 1995. Neural Networks in Computer Intelligence. McGraw-Hill, New York
Cao, S., Hermanrud, C., Lerche, I., 1988a. Estimation of formation temperature from
460 pp.
bottom-hole temperature measurements: COST #1 well, Norton Sound,
Manetti, G., 1973. Attainment of temperature equilibrium in holes during drilling.
Alaska. Geophysics 53, 1619–1621.
Geothermics 2, 94–100.
Cao, S., Lerche, I., Hermanrud, C., 1988b. Formation temperature estimation by
Morton, J.C., 2009. Boosting a fast neural network for supervised land cover
inversion of borehole measurements. Geophysics 53, 979–988.
classification. Computers & Geosciences 35, 1280–1295.
Davis, A.P., Michaelides, E.E., 2009. Geothermal power production from abandoned
Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning internal representa-
oil wells. Energy 34, 866–872.
tions by error propagation. Parallel Distributed Processing 1, 318–362.
Da-Xin, L., 1986. Non-linear fitting method of finding equilibrium temperature
Saito, S., Sakuma, S., Uchida, T., 1998. Drilling procedures, techniques and test
from BHT data. Geothermics 15, 657–664.
results for a 3.7 km deep, 500 1C exploration well, Kakkonda, Japan.
Demuth, H., Beale, M., 1998. Neural Network Toolbox for Matlab-User’s Guide
Version 3. The MathWorks Inc., Natrick, MA 742 pp. Geothermics 27, 573–590.
Dı́az-González, L., Santoyo, E., Reyes, J., 2008. Tres nuevos geotermómetros Santoyo, E., Garcı́a, A., Espinosa, G., Hernández, I., Santoyo, S., 2000. STATIC_TEMP:
mejorados de Na/K usando herramientas computacionales y geoquimiomé- a useful computer code for calculating static formation temperatures in
tricas: aplicación a la predicción de temperaturas de sistemas geotérmicos geothermal wells. Computers & Geosciences 26, 201–217.
(‘‘Three new improved geothermometers of Na/K using computational and Schoeppel, R.J., Gilarranz, S., 1966. Use of well log temperatures to evaluate
geochemometric tools: application to the prediction of temperatures in regional geothermal gradients. Journal of Petroleum Technology 18,
geothermal systems’’). Revista Mexicana de Ciencias Geológicas 25, 465–482. 667–673.
Dowdle, W.L., Cobb, W.M., 1975. Static formation temperature from well logs—an Serpen, G., Palabiyik, Y., Serpen, U., 2009. An artificial neural network model for
empirical method. Journal of Petroleum Technology 11, 1326–1330. Na/K geothermometer. In: Proceedings of the 34th Workshop on Geothermal
Espinosa, G., Garcia, A., Santoyo, E., Hernandez, I., 2001. TEMLOPI/V.2: a computer Reservoir Engineering, Stanford University, Stanford, USA, pp. 1–12.
program for estimation of fully transient temperatures in geothermal wells Shen, P.Y., Beck, A.E., 1986. Stabilization of bottomhole temperature with finite
during circulation and shut-in. Computers & Geosciences 27, 327–344. circulation time and fluid flow. Geophysical Journal of the Royal Astronomical
Espinosa-Paredes, G., Espinosa-Martinez, E.G., 2009. A feedback-based inverse heat Society 86, 63–90.
transfer method to estimate unperturbed temperatures in wellbores. Energy Spichak, V.V., 2006. Estimating temperature distributions in geothermal areas
Conversion and Management 50, 140–148. using a neuronet approach. Geothermics 35, 181–197.
Farshad, F.F., Garber, J.D., Lorde, J.L., 2000. Predicting temperature profiles in Spichak, V.V., Goidina, A.G., 2005. Temperature prediction in geothermal zones
producing oil wells using artificial neural networks. Engineering Computations from borehole measurements using neural networks. Physics of the Solid Earth
17, 735–754. 4, 844–852.
Fomin, S., Chugunov, V., Hashida, T., 2003. Analytical modelling of the formation Steingrimsson, B., Gudmundsson, A., 2006. Geothermal borehole investigations
temperature stabilization during the borehole shut-in period. Geophysical during and after driling. In: Proceedings at Workshop on Decision Makers on
Journal International 155, 469–478. Geothermal Projects in Central America, El Salvador, pp. 1–10.
Fomin, S., Hashida, T., Chugunov, V., Kuznetsov, A.V., 2005. A borehole temperature Verma, S.P., 2005. Estadı́stica Básica para el Manejo de Datos Experimentales con
during drilling in a fractured rock formation. International Journal of Heat and la Aplicación en la Geoquı́mica—Geoquimiometrı́a (Basic Statistics for
Mass Transfer 48, 385–394. Experimental Data Reduction with Application in Geochemistry—Geochemo-
González-Partida, E., Garcı́a-Gutiérrez, A., Torres-Rodrı́guez, V., 1997. Thermal and metrics). Universidad Nacional Autónoma de México, México, DF 186 pp.
petrologic study of the CH-A well from the Chipilapa-Ahuachapan geothermal Verma, S.P., Andaverde, J., Santoyo, E., 2006a. Application of the error propagation
area, El Salvador. Geothermics 26, 701–713. theory in estimates of static formation temperatures in geothermal and
Goutorbe, B., Lucazeau, F., Bonneville, A., 2006. Using neural networks to predict petroleum boreholes. Energy Conversion and Management 47, 3659–3671.
thermal conductivity from geophysical well logs. Geophysical Journal Inter- Verma, S.P., Andaverde, J., Santoyo, E., 2006b. Statistical evaluation of methods for
national 166, 115–125. the calculation of static formation temperatures in geothermal and oil wells
Hasan, A.R., Kabir, C.S., 1994. Static reservoir temperature determination from using an extension of the error propagation theory. Journal of Geochemical
transient data after mud circulation. Society of Petroleum Engineering, Drilling Exploration 89, 398–404.
& Completion 9, 17–24. Verma, S.P., Dı́az-González, L., González-Ramı́rez, R., 2009. Relative efficiency of
Hagan, M.T., Menhaj, M., 1994. Training feed forward networks with the single-outlier discordancy tests for processing geochemical data on reference
Marquardt algorithm. IEEE Transactions on Neural Networks 5, 989–993. materials. Geostandards and Geoanalytical Research 33, 29–49.
Hermanrud, C., Cao, S., Lerche, I., 1990. Estimates of virgin rock temperature Wisian, K.W., Blackwell, D.D., Bellani, S., Henfling, J.A., Normann, R.A., Lysne, P.C.,
derived from BHT measurements: bias and error. Geophysics 55, 924–931. Forster, A., Schrotter, J., 1998. Field comparison of conventional and new
Hernández-Pérez, J.A., Garcı́a-Alvarado, M.A., Trystram, G., Heyd, B., 2004. Neural technology temperature logging systems. Geothermics 27, 131–141.
networks for the heat and mass transfer prediction during drying of cassava Yilmaz, S., Demircioglu, C., Akin, S., 2002. Application of artificial neural networks
and mango. Food Science and Emerging Technologies 5, 57–64. to optimum bit selection. Computers & Geosciences 28, 261–269.
Hsieh, B.Z., Wang, C.W., Lin, Z.S., 2009. Estimation of formation strength index of Zhang, G., Patuwo, B.E., Hu, M.Y., 1998. Forecasting with artificial neural networks:
aquifer from neural networks. Computers & Geosciences 35, 1933–1939. the state of the art. International Journal of Forecasting 14, 35–62.

View publication stats