SELF-LEARNING RESERVOIR MANAGEMENT

SPE 84064

L. Saputelli1,2,3, M. Nikolaou2, and M. J. Economides2,3

1PDVSA, 2University of Houston, 3SPE member

PRH 034 - Formação de Engenheiros nas Áreas de Automação, Controle, e Instrumentação do Petróleo e Gás -
aciPG, participante do Programa de Recursos Humanos da ANP - Agência Nacional do Petróleo - para o setor
Petróleo e Gás - PRH - ANP/MME/MCT

October 21 2004; Florianopolis

1
 Agenda

• Motivation: The reservoir management challenge

– What is the Problem?,
– What have been done?
– What are the challenges?
• Problem Formulation
• The specific objectives and scope of this research
• Reservoir modeling and identification
• Model Predictive Control
• Self Learning Reservoir Management
• Conclusions
• The Way Forward

2
 Objective of this presentation

• To review current petroleum production

issues regarding real time decision making
and,
• To present the results of a continuous self-
learning optimization strategy to optimize
field-wide productivity while satisfying
reservoir physics, production and business
constraints.

3
Reservoir Management is about a careful orchestration of technology, people & resources

The Reservoir Management Challenge

Exploitation Plan
Well location & number
Recovery mechanism Drill,
Drill, build
build &
& Operate
Operate
Surface facilities
Well intervention Monitor
Monitor
Compression &
Production
Treatment Plants
Well & Facilities
Injection
Facilities

Establish
Establish or
or revise
revise
Optimum
Optimum Plan
Plan

Control
Control

Drainage
Area

Subsurface
Subsurface
Update
Update Model
Model
Characterization
Characterization

4
 Hydrocarbon production system suffering major technical problems

Motivation
Traditional Problems Current Approach Challenges
Complex & risky operations More front-end engineering More data for analysis and
(Drilling, Workover, Prod.) and knowledge sharing integration limitations.

Poor reservoir prediction & Integrated Characterization & Long-term studies, Ill-posed
production forecasting Modern visualization tools tools, simple or incomplete.

Limited resources: injection Multivariable optimization, Models are not linked among
volumes, facilities, people. reengineering. different layers

Unpredictability of events: Monitoring & control devices, Poor Justification, real time
gas or water, well damage. Beyond well measurements analysis in early stage.

Poor decision making ability Isolated optimization trials Decisions made only on few
to tune systems, thus, not with limited success. pieces. Lack of Integration
optimized operations between subsurface-surface

5
 To develop a field-wide continuous self-learning optimization decision engine

Research Specific Objectives

• Model based control system used to continuously

optimize three-phase fluid migration in a multi-layered
reservoir

• A data-driven model that is continuously updated

with collected production data.

• A self-learning and self-adaptive engine predicts the

best operating points of a hydrocarbon-producing
field, while integrating subsurface elements surface
facilities and constraints (business, safety, quality,
operability).

6
 Combination of petroleum reservoir physics and process control technologies

Research Framework

Data
Data Model
Model System
System Reservoir
Reservoir Bi-layer
Bi-layer Close-loop
Close-loop
Handling
Handling Building
Building Identification
Identification Performance
Performance Optimization
Optimization Control
Control

• Data handling
– Data acquisition, filtering, de-trending, outliers detection
• Model building and identification
– Gray box modeling: empirical reservoir modeling
– Partial least square impulse response, neural network and sub-space
• Reservoir performance prediction
– Real time Inflow performance and well restrictions
– Havlena-Odeh Material Balance
• Bi-layer optimization of operating parameters
– Reservoir best operating point based on the net present value optimization
– Regulatory downhole sleeves and wellhead choke controls
• Closed-loop control with history-matched numerical reservoir model
– Study of the system behavior in closed-loop

7
 Attempt to solve two significant reservoir management challenges

Problem Definition
Injector
Injector -- Producer
Producer Profile
Profile Mngt.
Mngt.
• Control undesired fluid production
• Exploit efficiently multilayer horizons
• Characterize inter-well relationship
• Maximize reserves and production
• Control from surface measurement
Field-Wide
Field-Wide Management
Management
• Optimization fluid production (< bottle-necks)
• Commingle multilayer reservoirs
• Minimize production costs
• Maximize reserves and production
• Control from surface measurement

Gas

Crudo

h1 k1 Agua

hn kn

8
 By collecting data, a digital image is used to make decisions

Traditional (Ideal) Integrated Management Approach

1
9 2
System:
Data
Implementation Reservoir, Wells &
Collection
Surface Facilities

Database
Business Parameters 3
6
Constraints Check Condition
Applications

Reservoir
7
Model
8 Decision
Optimization
Making Surface
Well 4
Model Model

Business
Model
Exploitation
Options
5

9
 Spatial distribution of pressure as a time function of saturation

Reservoir Modeling: Fluid Transport in Porous media

Multiphase Darcy’s Law

This realization is not used in this
research, since it requires the
krp K ⎛ g ⎞ knowledge of parameters that cannot
vp = − ∇
⎜ p p − ρ ∇Z ⎟ be directly measured
µp ⎝
p
gc ⎠
Pressure Laplacian as a function of the saturation change

∂c krp K ⎡ ⎛ g ⎞⎤ ∂ ⎛ φ S p ⎞
+ ∇ ⋅ ( cv ) = 0 ∇ ⋅ ⎢c ⎜ ∇ p p − ρ p ∇Z ⎟ ⎥ = ⎜
Continuity
⎜ ⎟
⎟
Equation
∂t µp ⎣ ⎝ gc ⎠ ⎦ ∂t ⎝ β p ⎠

∆yi zi+1 yi+1

M W A∆xφ S p β p φ S p
c= = =
VM A∆x βp
∆zi
Molar density in terms of Porous Volumes
yi-1
zi-1

10
 Oil, water and gas flow as linear functions of the drawdown pressure

Reservoir Modeling: Flow through Wellbore

Radial Diffusivity Equation
K ⎛ ∂ 2 p 1 ∂p ⎞ ∂p
⎜ 2 + ⎟=
φµ (c f + c) ⎝ ∂r r ∂r ⎠ ∂t re
General Solution given by Exponential Integral
qµ ⎛ φµct r 2 ⎞⎟ rs
p (r , t ) = pi − Ei ⎜⎜ ⎟⎟ h k
4πkh ⎝ 4kt ⎠⎟
⎜ rw
pe
Wellbore flow given by logarithmic approximation
qµ 4kt
pwf = pi − ln pwf
4πkh γφµct rw2
Proposed IPR for continuous monitoring
Steady-state Equation for the Undersaturated Oil-Flow
kkro h ( pe − pwf )
qok = a0 + a1 ⋅ pek + a2 ⋅ pwfk + a3 ⋅ ( pwfk )
2
qo ,b = o

141.2 Bo µo ⎡⎣ln (re rw ) + s ⎤⎦

qwk = b0 + b1 ⋅ pek + b2 ⋅ pwfk + b3 ⋅ ( pwfk )
2

Inflow Performance (IPR) for Saturated reservoirs

q = c0 + c1 ⋅ p + c2 ⋅ p + c3 ⋅ ( p )
2
pb ⋅ J ⎡ ⎛ pwf ⎞ ⎤
2
⎛ pwf ⎞
* k k k k
qo = qo ,b + ⎢1 − 0.2 ⎜ ⎟ − 0.8 ⎜ ⎟ ⎥
g e wf wf
1.8 ⎢ ⎝ pb ⎠ ⎝ pb ⎠ ⎥⎦
⎣

11
 Average Reservoir Pressure is a function of net mass production

Reservoir Modeling: Average Pressure Modeling

Material Balance Equation Expansion of Oil and Original dissolved gas, Eo
Net Underground = + Expansion of Gas Caps, Eg
Withdrawal, F + Reduction of Hydrocarbon Pore Volume, E fw
Simplification + Natural Water Influx, We

f ⎡⎣ p ( t ) ⎤⎦ = g ( N p , G p , W p , We )

⇒ p = a0 + a1 ∫ qo + a2 ∫ qw + a3 ∫ qg + a4 ∫ qwi
dp
⇒ = b1qo + b2 qw + b3 qg + b4 qwi
dt
⇒ 1
∆t ( k
p −p
k −1
) k
≈ c0 + c1 ⋅ p + c2 ⋅ pwf
k
1 + c3 ⋅ ( pwf 1 ) + c5 ⋅ pwf 2 + c6 ⋅ ( pwf 2 )
k 2 k k 2

Proposed Pressure Modeling for continuous monitoring

( p) = ( p) + c3 ⋅ ( p ) + c5 ⋅ ( p )
k k −1 2 2
+ c1 + c2 ⋅ p k
wf 1
k
wf 1 + c4 ⋅ p k
wf 2
k
wf 2

12
 Vertical flow as non-linear functions of flow rates

Reservoir Modeling: Flow Through Pipes

Mechanical Energy Equation Pressure (psia)

2
dp udu g 2 f f u dL 0 1000 2000 3000 4000 5000

+ + dz + + dWs = 0 0

ρ
1,000 STB/D @ 500 SCF/STB; WOR=1
gc gc gc D 2000 1,000 STB/D @ 1000 SCF/STB; WOR=1
1,000 STB/D @ 500 SCF/STB; WOR=0
1,000 STB/D @ 1000 SCF/STB; WOR=0
4000
Single-Phase Solution, Incompressible
ρ 2 f f u 2 dL

Depth (ft)
6000
g
∆p = p1 − p2 = ρ∆z + ∆u +
2
8000
gc 2 gc gc D
10000
Two-Phase Solution, Hagerdorn & Brown (1965)
∆ ( u 2 gc )
12000
2
dp fm& 2
=ρ+ +ρ
m 14000
144
dz ( 7.413 ×10 D ) ρ
10 5
∆z

Proposed Pressure Drop Modeling for Continuous Monitoring

(p ) = b q + b q + b q + b4 ( q ) + b5 ( q ) + b6 ( q )
k
k 2 k 2 k 2
− pth
k k k k
wf 1 o 2 w 3 g o w g

13
Well operating point given by the intersection of reservoir and tubing performance

Reservoir Modeling: Well Deliverability

Tubing Head Pressure

pwf , [psia]
Choke
Tubing
Performance
pwf , [psia]

q, [BPD]

pwf , [psia]

Fractional Flow of Phases q, [BPD]

Reservoir
Performance
Bottomhole Flowing Pressure

Reservoir Pressure
q, [BPD]

14
 Knowing input-output relationships, reservoir could seen as a process plant

Reservoir as a Process Control System Structure

Backpressure Reservoir Rock Heterogeneity
Ambient Temperature Measured Unmeasured Reservoir Fluid Distribution
Flow Restrictions Disturbances Disturbances Scheduling
Injection Fluid Restriction

Feed forward path

Manipulated Inputs
Flow Choke
Zone Control Unmeasured Outputs
ESP Speed Well flowing Pressure: pwf
Gas
Gas Lift Reservoir Pressure: pres
Crudo

Controller Solvent Injection Agua Reservoir Saturations: So, Sw

Water Injection Flow Impairment: S, Kr’s
Heat Injection Zone Multiphase Flow: qo, qw, gq
Gas Injection
Drainage Area: A

Tubing Head Pressure: pTHP

Feed back path Tubing Head Temperature: TTHT

Multiphase Flow: qo, qw, gq
Solid Production, Water Analysis

Measured Outputs

15
 Recursive self-adaptive identification

Reservoir Model Identification

Physical System
pk +1 , qk +1
Set point Reservoir
Reservoir
sp Simulator
Simulator
p wf

qGsp Measure

Control
Control Model
Model
Control Interpret
Implementation
Implementation Identification
Identification
n+ N
qo ,opt
∑ hu
Optimize
yˆ k + n + j|k = i k + n + j −i + dˆk |k
Set point i = n +1
Reservoir Model
Reservoir Value
Value
Optimization
Optimization
⎡ 1 ⎤ LS Optimization Loop
+ a3 ⋅ ( p wf )
2
qok = a0 + a1 ⋅ pek + a2 ⋅ pwf
a3 ⎢⎢ p k ⎥⎥
k k
⎡q k ⎤ ⎡ a0 a1 a2 ⎤
⎢ o ⎥ ⎢ ⎥⎢ e ⎥ ˆ = Xθˆ + e
Y
q = b0 + b1 ⋅ p + b2 ⋅ p
k k k
+ b3 ⋅ ( p) k 2
⎢q k ⎥ = ⎢b b1 b2 b3 ⎥ ⎢ p k ⎥
w e wf wf ⎢ w ⎥ ⎢ 0 ⎥ ⎢ wf ⎥ ⎧N ⎫
⎢q ⎥ ⎢c min ⎨∑ ei2 ⎬ ⇒ ( X T X ) XT Y
-1
k
⎥⎦ ⎣⎢ 0 c3 ⎥⎦⎥ ⎢ k 2 ⎥
+ c3 ⋅ ( pwf )
2
⎢⎣ c1 c2
q gk = c0 + c1 ⋅ pek + c2 ⋅ pwf ⎢⎢( pwf ) ⎥⎥
k k g
⎣ ⎦
a ,b
⎩ i =1 ⎭
(
⎧qo, g , w = f1 p k , p k −1 ... qTk , qTk −1 ,...
⎪ )
⇔⎨
( k −1
⎪⎩ pres = f n p k , p k −1 ... qT , qT ,...
k
)

16
Example for Model Identification and Block Diagram

Empirical model whose

Identification
Identification
structure is determined
by first principles

Empirical
Empirical
Model
Model
d
qo
+ qw
pwf Reservoir
Reservoir
(Simulator)
(Simulator) qg

Outputs (Y)
Inputs (U) y1
Reservoir Pressure: P
qwinj qo y2
Producer Flowing Pressure, pwf1 u1 Oil Rate: qo
y3
Water Rate: qw
Injector Flowing Pressure, pwf2 u2 y4
Water Fraction: fw
y5
Gas Rate: qg
y6
Water Injection Rate: qwi

17
 Model Identification Experimental Set-up

Windows and Eclipse Environment

Generate Reservoir Run Convert

Run Run
Data Numerical Summary Eclipse
Eclipse Matlab
File Model File To Excel

Matlab Environment Level

Read
Subspace Excel
Identification File
x=0 A
Plot Rsc x,y σ =1 U, Y Ac , Ad
Neural Split Data Auto Select Input cumsum(A)
Calculated
Network Test & Pred Scale and Outputs diff(A)
& Measured

FIR PLS Rescale

Parameters

Least Squares
Estimator

18
Predictions Using Empirical Structured models

19
Errors Using Empirical models

20
Coefficients Using Empirical models

21
 MPC minimizes future prediction error while satisfying input constraints

Model Predictive Control

At time k future predictions of the output y can be made as
n+ N n+ N
yˆ k + n + j|k = ∑
i = n +1
hi uk + n + j −i + dˆk |k where dˆk |k = yk − ∑ hu
i = n +1
i k −i

Minimization Problem to solve

⎡ ⎧ p ⎫⎤
⎢ m in ⎨ ∑ ( yˆ k + n + j | k − y )
m
+ R ∑ ∆ u k2 + j − 1| k ⎬ ⎥
SP 2

⎢ ⎩ j =1 j =1 ⎭⎥
⎢ s .t . ⎥
⎢ ⎥
⎢ y m in ≤ yˆ k + n + j | k ≤ y m a x j = 1, L , p ⎥
⎢ ⎥
⎢ u m in ≤ u k + j − 1| k ≤ u m a x j = 1, L , m ⎥
⎢u = u k + m − 1| k i = m ,L , p − 1 ⎥
⎢ k + i |k ⎥
⎢⎣ ⎥⎦

• Controls operation while optimizing performance

• Done over a receding or moving horizon

• Requires a setpoint from an upper level Set Point Tracking Example

All Variables normalized so that They have zero mean and Std. Dev = 1

22
 Example for Control and Block Diagram

PLS
PLS Impulse
Impulse
Identification
Identification

Empirical
Empirical
Model
Model
d
qo , sp
∆qo + qo
qw, sp + MPC
pwf
Reservoir
MPC Reservoir qw
qg , sp Controller
Controller (Simulator)
(Simulator)
-
qg

Outputs (Y)
Inputs (U) y1
qwinj qoT Reservoir Pressure: P
y2
Producer Flowing Pressure, pwf1 u1 Oil Rate Layer 1: qo1
y3
Oil Rate Layer 2: qo2
Producer Flowing Pressure, pwf2 u2 y4
Water Rate Layer 1: qw1
Layer 1, kh1 y5
Injection Flowing Rate, qwinj u3 Water Rate Layer 2: qw2
Layer 2, kh2 y6
Water Injection Rate: qwi

23
 MPC minimizes future prediction error by satisfying input constraints

Model Predictive Control Response

24
 Continuous self-learning optimization decision engine

New Self-learning Reservoir Management Technique

Physical System
pk +1 , qk +1
Set point Reservoir
Reservoir
sp Simulator
Simulator
p wf

qGsp Measure

Control
Control Model
Model
Control Interpret
Implementation
Implementation Identification
Identification
n+ N
qo ,opt
∑ hu
Optimize
yˆ k + n + j|k = i k + n + j −i + dˆk |k
Set point i = n +1
Reservoir Model
Reservoir Value
Value
Optimization
Optimization

LP Optimization Loop
⎧ ⎫
max ⎨ NPV = ∑ f ( qo , qw , qg ,$, ∆T ) ⎬
N

⎩ qo ,qw ,qg 1 ⎭
⎪⎧ pmin ≤ pk + p , k ≤ pmax
s.t. ⎨
⎪⎩qmin ≤ qk + p ≤ qmax

⇔ {qˆo ,opt , qˆ g ,opt , qˆw,opt

25
 Upper optimization layer passes the best operating point to lower layer

Multilayer Reservoir Control Model

Linear
Linear Programming
Programming Net
Net Present
Present Value
Value Reservoir
Reservoir Forecasts
Forecasts
Optimizer
Optimizer Function
Function

Optimization Layer

Regulatory Layer
Information
PLS
PLS Impulse
Impulse
Identification
Identification

Empirical
Empirical
Model
Model
d
qo , sp
∆qo + qo
qw, sp + MPC
pwf
Reservoir
MPC Reservoir qw
qg , sp Controller
Controller (Simulator)
(Simulator)
-
qg

26
 Best operating point (LP) problem subject to well constraints

Linear Optimization Problem

⎧ ⎫
max ⎨ NPV = ∑ f ( qo , qw , qg ,$, ∆T ) ⎬
N

⎩ qo ,qw ,qg 1 ⎭

⎡( qok Po + qgk Pg − qwp

k
C − q k
C ) ∆T − I k
− C k
⎤ (1 − r k )
⎣ ⎦
N
NPV = ∑
wp wi wi k T F
k ⋅∆Tk
k =1
(1 + i ) 365

VLP 4: fw,max +pTHP,min

Down hole Flowing Pressure , psia

Reservoir
Performan pthp
VLP 3: fw,max +pTHP,max
c e
4 fw
VLP 2: fw,min +pTHP,min
pwf,max 3
2 VLP 1: fw,min +pTHP,max
1

pwf,min

ql,min ql,max Total Liquid Rate, BPD

27
 The self-learning cased permitted less water and more oil produced

Injector-producer Management Problem Results

Experimental Base: History-matched Model from El Furrial, HPHT, deep onshore, light oil

Base Case No control Self Learning Case

• Early water irruption reduced • Water irruption detected and controlled
• High water cut reduced well’s vertical lift • Zone shut off permitted better well’s vertical lift
• Further recovery possible at a greater cost • Recovery accelerated at a minimum cost

28
 Clear benefits from extra little oil but with a lot less effort.

Field-wide life cycle comparison Results

Oil rate Oil Cumulative

Self-Learning 5%

∆Np=500 Mbbls

∆Rev=$5 Million Non-Controlled

Self-Learning

Non-Controlled

Water rate

Wp, Produced Water Cumulative

ntrolled
∆Wp= -18 MMbbls Winj, Injected Water Cumulative Winj Non co
Non-Controlled
∆Wi= -19 MMbbls
-55%
∆Rev= -$92.5 Million led
trol
on con
N
Wp
Winj Controlled
-78%

Self-Learning Wp Controlled

29
 Continuous self-learning optimization decision engine

New Self-learning Reservoir Management Technique

Physical System
pk +1 , qk +1
Set point Reservoir
Reservoir
sp Simulator
Simulator
p wf

qGsp Measure

Control
Control Model
Model
Control Interpret
Implementation
Implementation Identification
Identification
n+ N
qo ,opt
∑ hu
Optimize
yˆ k + n + j|k = i k + n + j −i + dˆk |k
Set point i = n +1
Reservoir Model
Reservoir Value
Value
Optimization
Optimization

QP Optimization Loop LP Optimization Loop LS Optimization Loop

⎧ p ⎫ ⎧ ⎫ ˆ = Xθˆ + e
Y
max ⎨ NPV = ∑ f ( qo , qw , qg ,$, ∆T ) ⎬
N
min ⎨∑ ( yˆ k + j − y SP ) + R ∑ ∆uk2+ j ⎬
m
2

∆u
⎩ j =1 ⎭ ⎩ qo ,qw ,qg ⎭ ⎧ ∞ 2⎫
min ⎨∑ ei ⎬ ⇒ ( XT X ) XY
j =1 1 -1

s.t. ⎧⎪ pmin ≤ pk + p , k ≤ pmax a ,b

⎩ i =1 ⎭
s.t. ⎨
ymin ≤ yˆ k + j|k ≤ ymax ; j = [1, p ] ⎪⎩qmin ≤ qk + p ≤ qmax (
⎧qo , g , w = f1 p k , p k −1 ... qTk , qTk −1 ,... )
⎪
umin ≤ uk + j|k ≤ umax ; j = [1, m] ⇔⎨
( k −1
⎪⎩ pres = f n p k , p k −1 ... qT , qT ,...
k
)
uk +i|k = uk + m −1|k ; i = [ m, p ] ⇔ {qˆo ,opt , qˆ g ,opt , qˆ w,opt

30
 Summary & Conclusions

• Novel multilevel self adaptive reservoir performance

optimization architecture
– Upper level calculates the optimum operating point
– Based on NPV
– Optimum set point passed to underlying level
• Feasibility of the method demonstrated through a
case study
– Reservoir performance continuously optimized by an
adaptive self-learning decision engine
– Method capitalizes on available remotely actuated devices
• Algorithm feasible for downhole implementation
– Impart intelligent to downhole and surface actuation devices

31
 Acknowledgement

• Research work was done under the guidance of Dr.

Michael J. Economides and Dr. Michael Nikolaou at
the University of Houston

• Research partially funded by PDVSA and Cullen

College of Engineering Research Foundation at the
University of Houston

• Academic access to software technology: EPS,

Stonebond Technologies, KBR’s Advanced Process
Control framework.

32