Professional Documents
Culture Documents
Load Sharing Optimization of Parallel Compressors
Load Sharing Optimization of Parallel Compressors
AbstractThis paper deals with the problem of optimizing turbines on others, determining the best choice for the
the load distribution and the on/off switching sequences of number of operating units and the speed of the
parallel compressor units in natural gas pipeline compression corresponding drivers leads to a complex optimization
stations. Natural gas pipelines are used to deliver gas from problem. The main objective is to minimize the energy
production sources to customers. Compression stations on these consumption for a given compression station set point.
pipelines are generally composed of the interconnection of Constraints arise from the operational requirements of the
several compressors units and the aim of the load sharing
compressors such as limits on the rotating speeds and safety
optimization is to operate these units in an energy efficient way
while continuously satisfying the varying demand of gas flow. criteria such as avoiding of surge and stonewall conditions.
As the gas demand changes start-up and shut-down of Boosting stations on natural gas pipelines typically consist
compressor units might be required and the impact of these of several compressor trains in parallel where each train is
switching events on the expected lifetime of the compressors composed of a compressor unit, control valves, tanks, and
also needs to be taken into consideration. Two types of coolers as illustrated in Fig. 1. Therefore, the decision
operational constraints are involved: continuous constraints variables in such systems are the set of active compressors in
concerning the conservation equations of mass and flow and the parallel arrangement and their respective speed of
combinatorial constraints concerning the possibility of rotation.
changing the number of active compressors. Overall the
optimization problem can be formulated as a mixed integer Process and
nonlinear programming (MINLP) problem. In this article the Safety Valves
ms
4060
The compressor map also provides information on while in case of a gas turbine driven compressors, assuming
operational limits. On the left side of the map a boundary a GT efficiency K GT and a lower heating value LHV , the
called the surge line is defined; this line represents the limit fuel consumption is given by the following equation as
on the minimum flow that can be elaborated by the described in detail in [3]:
compressor for a given head or pressure ratio. If this limit is Pow shaft
exceeded then a flow instability called surge occurs, which m f (5)
can cause thermal and mechanical stress to compressor LHV K GT
blades potentially leading to damages and eventually also to The compressor station operating point is defined by
machine failure. On the right side of the map another limit using a static model of the discharge load, which has been
(called choke or stonewall) is underlined; it is the line which modeled as a gas pipeline based on partial differential
corresponds to the maximum flow that can be reached by the equations for flow and pressure taken from [4]. The pipeline
compressor depending on the aerodynamic characteristics of model is linked to the rest of the model equations through
the discharge piping. At the top and at the bottom of the map the pressure loss as described by equation (6), where Pout is
two mechanical limits are defined: the maximum operating
speed (MOS) and the minimum operating speed (mos). The the compression discharge pressure and Pdem is the pressure
region within the described limits represents the subset of all at the end of the pipeline, for instance the pressure inside a
feasible operating points for the compressor in examination. gas storage tank near a city.
Starting from experimental data or estimates of the Pout Pdem O P0T m tot
2
Z ( Pout )
manufacturer i.e. the compressor map speed and efficiency (6)
L 2 DA Z 0T0 U 0 Pout
2
can be defined for every feasible operating point by means
of polynomial approximations. To fit the speed and The compression discharge pressure is given by a spatial
efficiency data, a second degree polynomial was used: discretization of the momentum equation across the length
of the pipe, as presented in [4]. The pressure loss is mainly
z D00 D10m
c D01 pratio D20m c 2 D11m c pratio D02 pratio2 (1) due to friction losses which are a quadratic function of the
total mass flow. Summarizing, the optimization problem has
been formulated by introducing the static model of the
where z can be either the compressor speed Z or the compressor station focused on the energy consumption and
efficiency K p of the working point and D ij are the the static model of the pipeline used for defining the
polynomial coefficients. Eq. 2 represents the role of the discharge conditions. It should be noted that the partial
efficiency in the calculation of the shaft power as considered differential equations for the pipeline are used only once to
in [3]: determine the system resistance for a given total flow. It is
not necessary to solve them at every iteration of the
yp optimization problem.
Pow shaft m c ( 2)
Kp
III. OPTIMIZATION ALGORITHM AND PARAMETER
IDENTIFICATION
where y p is the polytropic head, given by:
A. OPTIMIZATION ALGORITHM
This problem can be classified as an MINLP problem
n 1
( v ) since it contains nonlinear functions in both the objective
Z in RT in n v P2 nv
yp 1 ( 3) function and the constraints and also combinatorial aspects
MW n v 1 P1 [4]. The general form of an MINLP can be stated as follows:
min f ( x, y)
while K p and nv are the polytropic efficiency and polytropic
s.t.
exponent, respectively. MW is the molecular weight of the P
g ( x , y) d 0
gas mixture, Zin the inlet compressibility factor, m c the mass x X Z n , y Y
flow through the compressor, Tin the suction temperature,
P1 the suction pressure and P2 the discharge pressure. In where x and y are the decision variables of the optimization
n p
case of compressors driven by electrical motors, assuming problem, X and Y are polyhedral subsets of R and R
an efficiency KVSD , the power consumption is given by: respectively. The functions f : X u Y o R and
g : X u Y o R m represent the cost function and constraints
Powshaft of the optimization problem. The optimization problem
Powelectical ( 4)
KVSD solver used in this case study is BONMIN which is an open
4061
source code for solving general MINLP problems. More process controller in a suboptimal way. In the following the
information on BONMIN and on the interface used in this online estimation of the map parameters is discussed. Since
study can be found in [8] and [9]. There are several speed and efficiency can be approximated by polynomials,
algorithms that can be used with BONMIN; for the present the problem of estimating the coefficients is a linear
case the approach based on branch and bound (BB) was regression problem, which can be solved via least squares
considered. In the particular arrangement used in this case methods [11].
study the nonlinear solver IPOPT is used for the solution of Defining the compressor characteristics as:
the continuous sub problems and MUMPS for solving the
linear system of equations. A reference of IPOPT can be U(E) = (E) (9)
found in [10]. The optimization formulation contains in where - and are vectors representing the polynomial
addition to the energy costs also costs for switching on or off coefficients and the actual measurements respectively, y is a
the compressors. For a given operating point the switching measure of speed or efficiency depending on which set of
costs help to avoid the starting up of a new compressor until polynomial coefficients are being estimated, while , U is its
a certain energy saving threshold is exceeded. The switching
costs included in the objective function are given by: estimated value. T 0 represents the initial guess for the
parameters and S 0 its relative weight. The solution of the
C START _ UP yi 1 xi C SHUT _ DOWN xi 1 yi (7)
error minimization problem is given by:
where yi is a decision variable that indicates if the 1
compressor i should be switched on or off in the current N
N
operating condition and xi is the current on or off operating -N S0 I(i 1)I(i 1)T S0T0 I(i 1)y(i) (10)
i 1 i 1
state of the same compressor, while CSTART_UP and
CSHUT_DOWN are constants that represent the start-up and shut-
down costs respectively. In a similar way constraints
corresponding to active or inactive compressors are handled IV. RESULTS
by using the binary variable yi. As an example the surge A. STEADY STATE OPTIMIZATION
constraint is modified as follows:
In the following, the optimization results for the case of a
s1m c ,i s0 (1 yi ) Psurge t pratio (8) gas boosting station composed of five compressors driven by
variable speed electrical motors and discharging into a 100
where s1 and s0 define the a surge control line depending km long pipe with a diameter of 1.7 m, are presented. Table
1 summarizes the main parameters used for the simulations.
c,i and pressure ratio pratio over
on the compressor flow m The problem size for the case study was x Z 5 and y R 5
the compressor map, while Psurge is an appropriate constant for the decision variables with 35 nonlinear constraints, 37
value used to satisfy the inequality also when the compressor linear constraints and 60 bound constraints.
has not been selected to operate. TABLE I
Finally, a note on the performance of the solver used is in MAIN PARAMETERS USED IN THE SIMULATION CASE STUDY
order. BONMIN is a local solver and thus it cannot Suction Pressure (bar) 13.2
guarantee global optimality of the solution. To counter this Suction Temperature (C) 59.3
limitation, a warm starting technique is utilized along with Nominal Weight Flow Wet (kg/sec) 68.2
an approach where the optimization problem is solved using Molecular Weight (Kg/Kmol) 19.62
various starting points randomly distributed over the search Nominal kW Required (at compressor coupling) 13150
space. This method results in a better local solution even for Nominal Polytropic Efficiency (%) 85%
a small set of starting points. Typical execution times with 5 Rated Speed (rpm) 8370
to 10 initial guesses were 90.5 to 177 seconds on an Intel Maximum Speed (rpm) 8789
Core i7 processor at 2.2 GHz running on Windows 7.
Speed and efficiency polynomials for one of the
B. PARAMETER ESTIMATION compressors are illustrated in Fig. 5. The optimization
results are presented in Fig. 6 as a schedule for the full range
As introduced in Eq. 1, the optimization uses fitted
of station mass flows that the compressor station can cover.
approximations for compressor maps and efficiency maps.
In this figure, the individual mass flows are shown on the y
These approximations can become inaccurate due to
axis as a function of the total station mass flow on the x axis.
variations in gas properties, disturbances, equipment
These plots indicate that the optimizer recognizes which
damage, fouling or decreased life-time of the equipment.
compressors are more efficient for a certain interval of the
This can be avoided by online-parameter estimation using
total mass flow. In Fig. 7 the total power consumption of the
the available measurements for each individual compressor.
boosting station given by the optimization strategy is
Without online estimation, the optimization would lead to
compared with the one achievable with the equal load
steady-state errors which would have to be corrected by the
distribution approach. In the simulation for this traditional
4062
method the compressors have been activated in the same of saved power with the MINLP approach can reach up to 16
sequence given by the MINLP optimization, while the % compared to the conventional approach. It is also intuitive
activation point differs between the two methods. The equal that the room for power saving is limited when the station
load distribution approach brings a new compressor into operates at low flows (most efficient compressor at low flow
operation when the efficiencies of the actual operating running alone with/without recycle) and large flows (all
compressors fall below a predefined threshold, while with compressors running close to MOS without recycle).
the MINLP optimizer calculates the actual optimal point to
activate a new compressor. 90
Equal Dividing
80 Optimized
120 70
100
60
80
50
60
40
40 30
3 80
2.5 70 20
60
2 50
1.5 40
Pressure Ratio 10
Mass Flow
0
0 50 100 150 200 250 300 350 400
Station flow [kg/s]
1
flows calculated with the MINLP algorithm and with the equal
0.6
load distribution approach.
0.4
18
0.2 16
3 80 14
2.5 70
60
2 50 12
1.5 40
Power Saving (%)
4
mc 5 [kg/s] mc 4 [kg/s] mc 3 [kg/s] mc 2 [kg/s] mc 1 [kg/s]
100 2
50
0
50 100 150 200 250 300 350 0
100 -2
0 50 100 150 200 250 300 350 400
50
0 Station Flow [kg/sec]
50 100 150 200 250 300 350
Figure 8. Percentage of power savings at a given station flow with
100 the MINLP solutions with respect to the equal load distribution
50
0 approach.
50 100 150 200 250 300 350
B. DYNAMIC SIMULATION - PARAMETER
100
50 ESTIMATION AND CONTROL
0
50 100 150 200 250 300 350
As mentioned earlier, the load sharing optimization
100 procedure assumes that reliable models are available which
50
0 describe the performance of the compressors. When this
50 100 150 200 250 300 350
Station flow [kg/s] simplified assumption does not hold (i.e. due to fitting errors
Figure 6. Scheduling of the individual mass flow of or due to the fact that the original performance maps are no
compressors as a function of the total station flow determined longer valid), the optimization procedure can lead to the
by the solution of the MINLP problem. computation of compressor speed values that are suboptimal
thus providing an inaccurate flow distribution and mismatch
The comparison of the total power consumption given by to the desired station flow. This mismatch is eliminated by
the two approaches shows that the MINLP solution grants using a process controller which corrects the errors between
significant savings in energy consumption. As shown in Fig. the measured and desired station flows. In the simulations
8, for the range of total mass flow considered, the percentage
4063
2.2
such a controller is implemented along with a parameter 19
19 50
estimation procedure that estimates the parameters 18 0
50
0
2
describing the compressor maps of efficiency and speed. The 18
00
17
controller then uses the deviation in flows to correct the
19 0
50
19 50
50
18 0
1.8
0
17
18 0
optimized speeds to achieve the desired station flow. On the 00
Pressure Ratio
19 5 0
0
17
16
50 17 0
19 0
5
other hand, as the parameters of the compressor maps
0
16 0
18 50
1.6 0 16
15 0 50
18 00
become more accurate, the corrective action of the process 15 50
17 50
00
controller becomes negligible. Fig. 9 shows a schematic of
19 5
1 4 400 0
1.4
19 0
1 3 5 0 1 2 50
50
170
185
16 0
0
1 30
the proposed load-sharing and map-adaptation procedure.
18 0
0
15 5
1
17 50
0
0
16 50
15 0
0
0
14 0 0
17 00
1.2
14 0
16 00
0
13 00
50
Plant Set Points
1
3
5
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Map m /mc [kg/s]
Optimization c max
Adaptation 2
Updated
Figure 11. A comparison between the real and the estimated speed
coefficients
Optimal
Inputs
Suction pressure ( )
Control V. CONCLUSION
Discharge pressure (
Corrected torques (
Anti-Surge
Inlet flow (
A framework for the optimization of parallel compressor
Control
operation in gas compression stations is presented. The
proposed strategy is based on formulating and solving an
Compressor Plant
MINLP problem, leading to optimal on/off switching
Figure 9. Proposed load-sharing-optimization map-adaptation sequences of the units and the corresponding optimal load
solution for parallel compressor boosting stations. distribution. Compared to conventional approaches, the
The load-sharing-optimization parameter-estimation/map- proposed strategy leads to significant energy savings. Since
adaptation procedure has been tested in a dynamic the performance characteristics of the compressors deviate
simulation by introducing a mismatch in the second from their normal values due to wear, fouling or other
compressor between the real compressor characteristics and factors, a procedure is presented for estimating the actual
the one used by the optimization and estimation routines. A map parameters online. The updated maps are then used to
trajectory profile of the required total mass flow has been change the model used in the optimization algorithm.
followed and the routine for the parameter identification has
REFERENCES
been applied to the acquired data. In Fig. 10 and 11 the real
maps (color lines) are compared with the estimated maps [1] H. Devold, Oil and Gas Production Handbook, ABB ATPA Oil and
Gas, 2006.
(red lines) of efficiency and speed curves. The trajectory [2] H. H. Nguyen, V. Uraikul, C.W. Chan and P. Tontiwachwuthikul, A
imposed on the maps represents the operating points comparison of automation techniques for optimization of compressor
followed by the compressor during the simulation. It is scheduling, Advances in Eng. Software, Vol. 39 Issue 3, 2008.
[3] M. Abbaspour, K. S. Chapman, P. Krishnaswami, Nonisothermal
important to highlight that the estimation is most accurate in Compressor Station Optimization, Journal of Energy Resources
the areas of the compressor map that have been explored. Technology, vol. 127, issue 2, 2005.
This is expected since the parameter estimation is (locally) [4] S. Moritz,A Mixed Integer Approach for the Transient Case of Gas
Network Optimization, PhD thesis, Darmstadt University, 2007.
valid near the point of approximation. [5] C. Floudas, Nonlinear and mixed-integer optimization.
2.2
0. 6 Fundamentals and applications, Oxford University Press, 1995.
5
[6] A. Cortinovis, D. Pareschi, M. Mercangz, and T. Besselmann
2
0.
Model Predictive Anti-Surge Control of Centrifugal Compressors
0.
5
Automatic Control in Offshore Oil and Gas Production, Norwegian
0.6
0.7
0.65
1.8
5
0. 85
Elsevier, 1978.
.6
00. 6
5 0. 7
75
6.65
5
0.8
1.2 0.8 programs, Discrete Optimization, vol. 5, no. 2, May 2008.
0 .8 75 0.7 6.6 [9] J. Currie and D. I. Wilson, OPTI: Lowering the Barrier Between
0.75 0.7 00.
0. .65
0
Open Source Optimizers and the Industrial MATLAB User,
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Foundations of Computer-Aided Process Operations, USA, 2012.
m /mc
c
[kg/s]
max
[10] A. Wchter and L. T. Biegler, On the implementation of an interior-
2
point filter line-search algorithm for large-scale nonlinear
Figure 10. A comparison between the real and the estimated programming, Mathematical Programming, vol. 106, no. 1, 2006.
efficiency map at the end of a simulation run covering an [11] S. Bittanti, Model estimation and adaptive systems, Pitagora
operating range indicated by the solid trajectory. Editrice, Bologna, first ed. 1990.
4064