Professional Documents
Culture Documents
Reservoir Simulations
Pengju Wang, SPE, and Michael Litvak, SPE, BP
ferent iterations and may lead to convergence difficulties for a 0 ⱕ xជi ⱕ 1, i = 1, . . . , nw, . . . . . . . . . . . . . . . . . . . . . . . . . . . . (3c)
reservoir simulation.
This paper addresses the first and second problem by present- where nw is the number of flow streams; xជi denotes the decision
ing a simple, yet effective, gas lift optimization algorithm. This variable for this problem; Qg is the total gas-flow-rate capacity of
paper also addresses the second problem by introducing a multi- the field; and qo,i, qg,i, and qlg,i are the oil, formation-gas, and
objective optimization formulation to minimize the lift-gas-rate lift-gas rates for Well i, respectively. In the optimal solution, xi⳱0
oscillations while maximizing the oil production. indicates that Well i should be shut in; xi=1 indicates that Well i
should produce at rates qo,i, qg,i, and qlg,i; and xi ∈ (0,1) indicates
Gas Lift Optimization Procedure that Well i should be choked back.
Overall Method. The gas lift optimization method developed in In this formulation, all the flow-rate constraints of the gas lift
this study takes into account flow interactions among wells optimization are satisfied. The optimal objective-function value of
through common surface pipelines. The method works as follows: Eq. 3 is regarded as the objective-function value corresponding to
1. Start with the existing lift-gas rates for all wells on automatic a set of lift-gas rates q៝ lg specified in Steps 1, 2, and/or 4 of the
lift-gas allocation. Solve the multiphase-flow problem in the SPN. overall gas lift optimization method.
Build a linear-programming model (described later in the Con-
straint Handling section) to scale production and lift-gas rates to Discussion. The results generated by the new gas lift optimization
satisfy flow-rate and/or velocity constraints. Denote the objective procedure will be suboptimal because of the following two facts:
function value obtained in this step as f 0. 1. The function evaluation procedure employed in the overall
2. Select a well on automatic lift-gas-rate allocation, Well i. optimization algorithm uses an optimization procedure with sim-
Denote its lift-gas rate at this stage as q0lg,i. Increase its lift-gas rate plified assumptions. As a result, the function value obtained from
by␦qlg,i. Solve the multiphase-flow problem in the SPN with the this procedure is only an approximation.
updated lift-gas rates, and scale production and lift-gas rates to 2. The new method is a local search method and may get stuck
satisfy the flow-rate constraints. The value of the objective func- at a local suboptimal point. In several case studies, it was demon-
tion obtained in this step is f 1. strated that the new method produces quality results for long-term
3. Compute the gas lift efficiency for Well i using Eq. 1: reservoir-development studies.
emin is a parameter used to control how easily a lift-gas rate can
f1 − f0 escape from its current value. If emin is large, the lift-gas rate is not
e= . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1) sensitive to small changes in reservoir and operation conditions.
␦qlg,i Consequently, the result will be less near the true optimum, but
If eⱖemin, where emin is the user-specified minimum gas lift effi- there will be a lesser simulation-convergence problem from lift-
ciency coefficient, update f 0 by setting f 0⳱f 1, and go to Step 6 gas-rate oscillations. Conversely, if emin is small, the allocated
with the increased lift-gas rate for Well i. If 0ⱕeⱕemin, reset the lift-gas rates will be more noisy, but the solution will be closer to
lift-gas rate of Well i to q0lg,i and go to Step 6. If e<0, reset the the true optimum.
lift-gas rate of Well i to q0lg,i and go to Step 4. To facilitate later discussions, the above overall gas lift opti-
4. Decrease the lift-gas rate of Well i by ␦qlg,i, where ␦qlg,i>0. mization method will be referred to as the GLINC method.
Solve the multiphase-flow problem in the SPN with the updated
lift-gas rates. Optimally scale the production rates and lift gas rates Lift-Gas-Rate Damping
to satisfy flow-rate constraints. The value of the objective function For constrained gas lift optimization problems, there may exist
in this step is f 2. cases in which multiple vastly different lift-gas distributions result
5. Compute the gas lift efficiency for Well i with Eq. 2: in similar total-oil-rate increases. For such cases, although moving
from the current gas lift injection scenario to another gas lift in-
f2 − f0 jection scenario may increase the total oil production by only a
e= . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (2) small amount, the resulting production rates for individual wells
− ␦qlg,i
can be significantly different, making simulation convergence dif-
If eⱖemin, update f 0 by setting f 0⳱f 2, and go to Step 6 with the ficult to achieve. When gas lift injection scenarios oscillate fre-
decreased lift-gas rate for Well i. Otherwise, reset the lift-gas rate quently in different Newton iterations, the computational effi-
of Well i to q0lg,i. ciency of the simulation run deteriorates significantly. In the
guarantee local optimums. To verify the performance of the SP and because the SP and GLINC methods have parameters to control
GLINC method further, two conventional optimization methods the balance between injection costs and oil-rate increases, while
were applied to the constrained gas lift optimization problem. The the current implementations of the GA method and COBYLA
first method is a genetic algorithm (GA) (Holland 1975). In this method do not have this functionality. Figs. 4, 5, and 6 compare
method, the lift-gas rates for wells on automatic lift-gas allocation the lift-gas rate, oil rate, and gas/oil ratio (GOR) allocated by the
are selected as the decision variables and encoded as a binary four optimization methods for Well A2, respectively. The lift-gas
string (or genes). This method encodes multiple sets of solutions rates allocated by the four methods follow roughly the same trend.
(lift-gas-rate distributions) into populations, evaluates the fitness Although the absolute lift-gas-rate differences between the four
of each population (total oil rate for a given set of lift-gas rates methods are significant, oil rates and GORs are similar. For a gas
subject to flow-rate constraints), and evolves the populations lifted well, gas lift efficiency decreases to a small value as the
through a number of generations by means of selection, crossover, lift-gas rate increases beyond a certain value. In actual field op-
and mutation. As in the GLINC method, the flow-rate constraints erations, the wells will be operated with lift-gas rates similar to
of surface facilities are handled in the function evaluation proce- those obtained from the GLINC method, which are considered to
dure using the linear-programming model (Lo et al. 1995) de- be optimal in this study.
scribed in the section on Constraint Handling. In this example, the Efficiencies of the four optimization methods are compared in
population size was set to 10, and each GA optimization took Table 1. For all runs, gas lift optimizations were invoked only in
20 generations. the first three Newton iterations of a timestep. Because the SP and
The second method is a direct-search algorithm that tries to GLINC methods usually require much less CPU time for a single
maintain a regularly shaped simplex throughout the iterations. The instance of gas lift optimization than is required by the GA and
method used in this study was implemented by Powell (1992) and COBYLA methods, the timestep controls of the four runs were
is referred to in this paper as the COBYLA method. As in the GA handled differently to optimize the total CPU time required for
method, the lift-gas rate for each well on automatic lift-gas allo- individual runs. The runs with the SP and GLINC methods had a
cation is selected as the decision variable, and the flow-rate con- maximum timestep size of 6 days, while the runs with the GA and
straints of surface facilities are handled in the function evaluation COBYLA methods had a maximum timestep size of 10 days,
procedure using the linear-programming model (Lo et al. 1995). resulting in a lower total number of timesteps and, thus, a lower
Results from the four different rate-allocation methods are number of gas lift optimization calls than the same runs with a
shown in Figs. 2 through 6. Although the field cumulative oil maximum timestep of 6 days would have had. Under these set-
productions from the four optimization methods are close (Fig. 2), tings, Table 1 shows that runs with the SP and GLINC methods
the daily field lift-gas-injection volumes from the four methods are required significantly less CPU time on both well management and
significantly different (Fig. 3). The SP and GLINC methods allo- overall simulation than the runs with the GA and COBYLA meth-
cated significantly less lift gas than the GA and COBYLA methods ods required.
Fig. 4—Field Example 1: normalized lift gas rate for Well A2 Fig. 5—Field Example 1: normalized oil rate for Well A2 allo-
allocated by the four rate-allocation methods. cated by the four rate-allocation methods.
Fig. 7—Field Example 2, Case 1: normalized daily lift-gas rate Fig. 8—Field Example 2, Case 1: normalized field cumulative oil
for Well A3 obtained from the SP method with and with- production obtained from the SP method with and with-
out damping. out damping.
Fig. 9—Field Example 2, Case 2: normalized daily lift-gas rate Fig. 10—Field Example 2, Case 2: normalized field cumulative
for Well A4 obtained from the SP method with and with- oil production obtained from the SP method with and with-
out damping. out damping.