I 
GEOL. C ROAT. 49/2 237 - 242 6 Figs. 5 Tabs. ZAGREB 1996 

Sciellfljh: paper 

Production Characteristics and Reservoir Quality at the Ivanic 
Oil Field (Croatia) Predicted by Machine Learning System 

Zvonilllir HERNITZ J
• Mi ro DUREKOVIC' and Jos ip CRNICKJ 3 

PROCEEDINGS 

KCJf words: IV3 1l ic oi l fie ld, Reservoir qual ity, Predic­
tion o r oil production, Machine lea rnin g sys tem , 
Expert systems. 

Abstract 
AI [ile Ivallie oi l ficld, hyd rocarbons are accumulated in fine to 

medium gl"ilincd li(hareni lcs or the [van ie-Grad Format ion (Iva-sand­
sIOI1C~ member) of Upper Miocene age. Reservo ir rock s are divided 
inlo eight depositional (prod uc ti on) units (i l- ivlI l)' Deposit s o f each 
lInit arc c haracterized by th e ir own reservoir quaIil Y parameters 
(porosity, horizontal permeability, net pay ... ). Prod uction characteris­
tics of 30 wel ls have been studied by a simple slatistica[ method. Two 
major production wel[ categorics ("good producers" and " bad produc­
ers") have been found . The contribut ion of each deposi tional unit to 
th e 10lal produc lion of iln individual well was sl udied by using :\ 
machine IC:ll"ll ing syslem which can serve as <I n expert system she ll. 
The results have shown Ihal: 

[) the mOSt important factors affecti ng the total hydrocarbon produc­
tion of each well arc horizontal penneability and net pay thickness 
of the depos itional units i V! i lv and i lll , 

2) afler 20 wells had bee ll dri ll ed during the primary recovery peri od 
il was possi ble 10 predict the production category of a well to be 
drilled in the future with a rel iability of 80%. 

1. INTRODUCTION 

The Ivanie oil fi e ld is s illlated in the north-wes lern 

part or the Republic of Croatia (Fi g. I) . The fi e ld was 

d iscovered in 1962 on d rilling the Iva-2 well. 

Hyd roca rbons ha ve been accul1lulated in f ine 10 

mediulll -grained Iithare nites (Tvanic-Grad Formation, 

Iva-sa nd stones membe r) o f Upper Mi ocene age. Th e 

trap is o f stfuclUral orig in (an an tic lin e s triking from 

north -west to sou th -cast). T he s truc ture top de pth 

ranges from -1500 III to -1 600 m. The first phase of 
reservoi r developmenl was com pleted in 1966. Hyd ro­

carbon rese rves include oil, d issolved gas and a sma ll 
amOllnt of gas in a gas cap. The fie ld mainly produces 

under a dissolved-gas dri ve. The primary recovery pe ri ­
od has bee n characte ri zed by a rapid dccrcase of reser-

voir pressure and by an increase of GOR. In 1972, the 
sccondary recovery pe riod (waterfloodi ng) started . 

According to the genctic stratigraphic sequence con­
cept (GALLOW A Y. 1989) rese rvo ir rocks have been 
di vided into eight depositional events (unit s) named it -

iVJJJ (DUREKOVIC, 1995; Fig. 2). 
The Ivan ie oi l field produces from depositional units 

ilil-ivtl. Stratigraphi c and depositi onal analysis, pe r­
formed by apply ing the total and net pay thickness, per­
meabili ty and porosity c riteria, have shown that reser­
voi r rocks are of bette r quality in the south-west part o f 
thc fi e ld. An elongated zone of bctte r reservoir qua lity, 
stri kes from the no rt h-west to sOllth-east and co inc id es 
with the palaeod rai nage patt ern (DUREKOVIC. 1995). 

We ll production rate is re latcd to Ihe rese rvoir qua li ­
ty. Generall y, the total wcl l produc tion rate ca n be 
dc fin ed as a func tion o f individual depositional unit 

production. Dcposits of each unit have been character­
ized by their own reservoir quality parameters (lola I and 
net pay th ickness, permeability , oi l satu ration ... ) . From 
a rcservo ir q uality poi nt o f v iew, the purpose o f our 
work is to define the impo rlance of each depositional 
unit and its contribution to the to tal well produc tion 
appl ying an expert sys tem. Furtherm ore, wc int end to 
tes t the role of a machine learning sys tem in prcdic ting 
well produc tion charactc ristics. 

2. METHODS AND RESULTS 

AI the firs t s tage o f our work we have ana lyzcd thc 
average production behavior of 30 we ll s during the pri ­
mary and seco ndary recovery pe riod . A Producfion 

illdex (PR IND) has been calcu lated for each wel l tak ing 
into accou nt well production rates and production peri­
od duration (time). A s im plc sta tis tica l mc thod showed 
the exis tence of two prod uc ti on we ll ca tcgories Cf able 
I ). 

T he firs t catego ry ("bad produ cers") includes 13 
wclls (out of 30) . P rod uct ion index values range from 
1.5 to 5.5 during the pri mary recovery period and from 
1.5 to 7.4 in the secondary phase of developmcn t. The 

I Univer~i ! y of Zagreb, Facul ty of Mining, Geology & Petroleum Engineering, Picrol!ijeva 6, r.O.Box ) 86. HR-tOOOO Zagreb, Croatia . 

~ l NA-Naftapl in , Subiceva 29, II R~ [0000 Z:lgrcb, Croatia. 

J fr:lll:\ Gotovca 13. HR- [(I()) 0 Zagreb, Croatia . 



2}& 

seAL E 
o !IO l001<m _____ J 

Fig. I Locmion map. 

second category ("good producers") includes 17 well s. 
PR IND values vary fro m 4.8 to 12.2 in the prim ary 
recovery period and fro m 2. 1 to 10.8 in the secondary 
recovery phase. T he main differences between those 
two categories can be determined from average annual 
production rates (Figs. 3 and 4; see also Fig. 5 fo r well 
locati ons). 

Four reservoir quality parameters (variabl es) have 
been analysed for each deposit ional (producti on) unit in 
each well: 

POR - effec tive porosity (%); 

HPER - horizontal permeability (x 1 0·' 11m' ); 

DEF - net pay thickness (m); 

• SO - oil sa tu ration (0/0) . 

Reservoi r qua li ty parameters (20 in tO lal) have been 
named afte r Ihe deposit ional unit number. For example, 
depos it ional un it iUt has the fo ll ow ing variables: 3POR, 
3f-IPER, 3DEF, 3S0 ; i,v has 4POR , 4HPER etc. 

The ana lysis also included three parameters com-
1110n for all depositional units in individual well s: 

• KNV - oil/water contact depth (m); 

• JK - name of depositional uni t where oil/water 
cont act has been established; 

• SUB - ex istence of injection well (first neigh­
bour) (Yes/No); 

All these va riables (23 in total ) have been analyzed 
by a machine learning system named "Assistant Profes­
siOlla l" (CESTNIK et aI., 1987). It can automati call y 
const ruct decision trees from examples, and it can serve 

o 

'00 

'" 

'50 

'" ,m , 

W 1128 e -

SP '" 0",,100 

I-

Itll 

W 
! i ' I! 

~ I 
!j 
I 

-

Geologia Cro:ttic:t 49n 

Deposi tional r 
E -Iog unit 

(parase- 11>"11> 
marker 0..0" 

quence) <1l 
11> 

Rp 
<:-

i VlII ~o 
I) 0'" 

(;;'" 
i VlI 

",0 

16 
Q;-'" 

i" 
c 

15 
0 

ro 
iv "0 

:tJ 
I, - CI 

CI 
.0: 

iN 

13 

i ll l 
C 
0 

ro 
I, - "0 

:tJ 
i ll 

CI 

2 
I, "-

i, 

~ Rs5 

rig.2 Depositional uni ts of genetic stratigr:tphic sequence IVA. 

as an ex pert sys tem she ll for the acquired know ledge 
base (NOVINC, 1992; C RNfCK I et a I. , 1994). The 
Ass istant Professional expert sys tem is a very fl exible 
computer package wh ich allows engineers 10 participate 
in the analys is by selecting the most important variables 
in decision Iree nodes (CRNI( Xf , 1989). 

During data analys is by the Assis tanl Profess iona l 
learning system, several ph ases were perform cd which 
approach the best possi ble prediction results. This paper 
di scusses four of them: 

I) Phase I - the learning system aUlOmat ically gener­
ates the rul es needed to classify the well in one of 
I \vo possible production categories. The system lIsed 
23 variables [or each of" 30 wells. A decis io n dia-

Primary recovcry period Secondary rccovcry period Number of wells % 
Catcgory 
(group) 

II 

Table I Production index. 

(1962·1972) 

PRIND 
average 
PRIND 

t.5 - 5.5 3.2 
4.8 - 12.2 8. 1 

(1972·1995) (total 30) 

PRIND 
aver-age 
PRIND 

t.5 - 7.4 3.7 13 43.3 
2. I . 10.8 5.5 17 56.7 

L 30 100.0 



Hernitz. Durekovic & Cmicki: Production Characteristics and Reservoir Quality ntthe Ivnnic Oil Field (Croatia) Predicted by ... 239 

Well no. 
32 

c 
0 

24 tl 
0 
~ o ~ 
0.. 0 16 c ~ 
0 .. € 

" ro 
u 

1 I 
- ~ 

' 1f~l~~ 
I 

l-m 
-

V ,\ -

19 

2 5 

3 6 

-- 18 22 
8 e 

~ 
~ 

I ~ ~~~ 'r / r\::: , tv< l&l I '~ "It I 
-- 4 19 

0 

1960 1970 1980 

Fig. 3 Hydrocarbon product ion rate - "bad producers". 

gram (Tab le 2) shows that depositional units iv, i,v 
and i lll affect the well production ratc much more 
than depositional units iv, and ivll' The most impor­
tant variables are horizontal permeability and net pay 
thickness (5I-IPER, 4DEF, 3DEF). Porosity (POR), 
well structura l position (JK) and depth of oil/water 
contact are less important, as well as the existence of 
an injection well (fi rst neighbour) in thc vicinity of 
(he producer (SUB). 

" 29.9 
4DEF 

" 10.5 > 10.5 

550 3DEF 
5: 49 .S >49.5 " 19.7 > 19.7 

JK 

I 

time 
1990 2000 

(year) 

2) Phase 2 - A defined number of wells (5, 8, 11 and 14 
from the total of 30) (Table 3) are selected by ran­
dom number generation process to be used for a 
decis ion diagram rel iability tes t. The n, the expert 
system was applied to the rest of the wells (non­
selected) taking into account acquired knowledge 
from phase 1 (including a decision diagram and the 
importance of variables [rom the tree). Finally, a 
new decision tree was generated and tested on a cho-

5HPER 

> 29.9 

3DEF 

" 19.7 > 19.7 

7DEF 
5: 2.5 > 2.5 

6POR 
::;; 4.5 >4.5 " 20.6 > 20.6 

Category 1 2 2 1 

Table 2 Decis ion diagram (phase I). 

32 
I I I 

I 

I I 
o 

1960 1970 1980 

2 2 

I 

1990 2000 

" 21.7 

1 

Well no. 

11 

13 

7 

12 

- 24 

27 

- 28 

29 

-- 15 

- 30 

time 

(year) 

2 

4POR 

> 21.7 

1 2 

Fig.4 Hydrocarbon produc­
tion rate - "good produ­
cers". 



240 

injector-well section 

p roduction group 
boundary 

oil/Woter 
contact 

c ontrolwetl 

• producers-group I 

• producers-group II 

°CI =======5;OO~ ....... 'OOOlm) 

Number of wells Number of corTeet Reliability 
learn ing testing answers (%) 

25 5 5 100.0 

22 8 7 85.5 
19 11 8 72.7 
16 14 10 7 1.4 

Table 3 Reliability of well category prediction (phase 2). 

Case (a) 
(without human intervention) 

NUlllber of well s 
number of reliability 

lear-Iling testing correct answers (%) 

25 5 3 60.0 
22 8 4 50.0 
19 II 5 45.0 
16 14 6 43.0 

• 

• 

• 

Geologia Croat ic:t 49/2 

Fig. 5 Production well 
category and locati on 
map. 

sen les t we ll s. The proced ure was repeated three 
times. Average results arc shown in Table 3. 

The gene ral conclusion is: pred ic tion reli abi lit y is 
proporti onal to the num ber of learning well s. In a ll 
cases rel iability is greater than 70%. 

3) Phase 3 - the ex pert system fo llowed the same proce­
dure as in phase 2, wi thou t laking into account 
knowledge from phase 1 or phase 2. Table 4 shows 
the re liability for two different cases. Tn case (aJ 

Case (b) 
(human intervention, fixed 

variables 5HPER, 4DEF, 3DEF) 
number' of rcli~)bility 

correct answers (%) 

3 60.0 
5 62.5 

6 55.0 
7 50.0 

Table 4 Rcliabilil y of 
well calcgory pre­
dicLion (phase 3). 



Ifernie/, Dureknvic & Crnick i: Production Characteristics and Reservoir Quality at the Ivanie Oil Field (Croalia) Predicted by ... 241 

. 25 control well 

\~ ~ contour line (m) 

production group 
boundary 

5 wells tested oftel ... 

D 
D 
V 
o 

step 1 -learning 
on 10 wells 

step 2-leorning 
on 15 wells 

step 3· learning 
on 20 wells 

step 4-leorning 
on 25 wells 

• • 

• 

_ true 

_ false 

stage of experiment 

• 

• 
I 
II 

initial·learning 
on 10 wells 

'bad producers' 

"good producers' 

• 

( 

_______ · 1540 

_____ ,156<) _-"'--~ _____ ·,580 ____ ~ 

o = l=~ .~ 
CI ====~ ____ ' '-~ Fig. 6 Succession of 

there was no human intervention in decis ion ana l y~ 

sis. In case (b) the first two levels (nodes), the most 
important variables were selected by us (human 
inte rvention ). Results show tha t human experi e nce 
can help the expert system to obtain beller result s. 

4) Phase 4 shows how the machine ieaming system can 
be used during reservoir exploration or/and deve lop­
ment. At the beginning we selected 10 wel! s o ut of 
30 according to thc drilling succession (well- I was 
drilled firs t, well-2 was drilled afler we ll- '--.). The 
sys te m generat es the deci s ion tree a utom at ic ally 
from the da ta base for those 10 wells. Re liability o f" 
the decision tree ana lysis was tes ted on fiv e wel ls 

Numbcr of wells Number of corrcct Reliability 
IC31"1ling testing answers (%) 

10 5 2 40.0 
15 5 3 60.0 
20 5 4 80.0 
25 5 4 80.0 

Table 5 Reliability of well category prediction (p hase 4). 

rese rvo ir explorati on 
orland development. 

(planned to be drilled) asslIming that we would be 
able to predict thc va lues or variab les needed for pre­

diction from the already known geolog ical situation. 
The procedure was repeated until 25 learning wells 
were reached. Results show (Table 5; see also well 
locations in Fig. 6) that 15-20 wells shou ld be drilled 
before the expert system can predict the production 
category of future (planned) we ll with a reliability of 
80%. 

3, CONCLUSIONS 

1) A machine learning syste m can be used in orde r to 
predict the s igni ficancc and contribution of individ­
ual production units, to the lOtal well produc tion tak­
ing iJ1lo account their reservoir quality (phase 1). 

2) Human intervention based on experience in many 
eases will help the expert system to obtain bel1er and 
more reliable results (phases 2 and 3). 

3) A machine learning system is 01" grea t he lp in plan­
ning reservoir exploration and/or dc ve lopment 
process by predicting well produc tion behaviour on 
the basis of the known geolog ical c harac te ri s tics of 



242 

reservoir rocks (phase 4). This gives us an opportu­
nity to predict the minimum number of wells needed 
to achieve a maximulll production effect. 

4) Further improvement could be achieved by a si milar 
analysis in 3D space. 

S) In the geological dec is ion maki ng process the 
machi ne learning systems are able to red uce the 
number of bad producer wells. It cou ld be recom ­
mended to usc the described or similar methods as it 
has importance in the oi l-production economy. 

4. REFERENCES 

CESTN IK , B., KONONENKO, I. & BRATKO, I. 
(1987): Assistant 86 - a knowledge - elicitation tool 
for sophi sticated uscrs.- In: BRATKO, I. & LA V­
RAC, N. (cds.): Progress in Machine Learning. Sig­
ma Press, 3 1-45, Wilmslow. 

CRN ICKI , J. (1989): Kno wledge based systems in 
Geology.- Proceedings of International Conference 

Geologia Croalica 49/2 

on Mathematical Methods in Geology, 2, 652-658, 
Pribram. 

CRNICKI, J., HERNITZ, Z. & CESTNIK, B. (1994): 
Oil forecasting in Slavonia (Croatia) by the Assis­
tant Professional Ex pe rt System ShcJl.- Croatian 
Geotechn ical Journal , 1/1-2,13-17, Varazdin. 

DUREKOVIC, M. (1995): Sedimentation and reservoi r 
c harac teri s tics of Iva sand stone at the oi l fi e ld of 
Ivanic.- Unpublished M.Sc. Thesis, Univers ily of 
Zagreb, 58 p., Zagreb. 

GALLOWAY , W.E. (1989): Genetic stratigraphic 
sequence in Bas in Analys is I: Architecture and gen­
esis of floo ding-surface bounded depositional 
Uoits.- AAPG Bulletin, 73/2, 125- 142. 

NOVINC, M. (1992): Possibilities fo r est imation of oil 
and gas in the eastern part of Drava bas in by a 
machine learning system.- Unpublished M.Sc. The­
sis, University of Zagreb, 85 p., Zagreb.