simulation & optimization of pressure swing...
TRANSCRIPT
Simulation & Optimization of Pressure Swing Adsorption
Yajun Wang & Alexander W. Dowling Lorenz T. Biegler
Department of Chemical Engineering Carnegie Mellon University
March 2014
Motivation
• Objective
-- Modify PSA simulation
-- Diagnose convergence problems in PSA optimization
• Pressure Swing Adsorption (PSA)
-- Gas separation (e.g. H2 – CO2)
-- In IGCC power plant
2 [1] Alexander W. Dowling, Sree R. R. Vetukuri, and Lorenz T. Biegler. Large-scale optimization strategies for pressure swing adsorption cycle synthesis.
Pressure Swing Adsorption (PSA) Superstructure[2]
State variables: CCO2 CH2 qCO2 qH2 T
5 operating steps (time slots): α β φ Pads Pdes
Additional control variables: ts Lbed
Cyclic steady state
3 [2] Anshul Agarwal. Advanced Strategies for Optimal Design and Operation of Pressure Swing Adsorption Processes. PhD thesis, Carnegie Mellon University, 2010.
PSA modeling[4]
• Mass balance 𝜖𝑏𝜕𝐶𝑖
𝜕𝑡+ 1 − 𝜖𝑏 𝜌𝑠
𝜕𝑞𝑖
𝜕𝑡+𝜕 𝑣𝐶𝑖
𝜕𝑥= 0
• Mass transfer 𝜕𝑞𝑖𝜕𝑡= 𝑘𝑖(𝑞𝑖
∗ − 𝑞𝑖)
• Energy balance
𝜖𝑡 𝐶𝑖𝑖
𝐶𝑝𝑔𝑖 − 𝑅 + 𝜌𝑠𝐶𝑝𝑠
𝜕𝑇
𝜕𝑡− 𝜌𝑠 ∆𝐻𝑖
𝑎𝑑𝑠
𝑖
𝜕𝑞𝑖𝜕𝑡+ 𝜕 𝑣ℎ
𝜕𝑥+ 𝑈𝐴 𝑇 − 𝑇𝑤 = 0
• Momentum balance
• Adsorption Isotherm
• Ideal gas equation
• Check valve equation, Bed connection equations, Auxiliary equations
PDEs
DAEs
4 [4] Ling Jiang, Lorenz T Biegler, and V Grant Fox. Simulation and optimization of pressure-swing adsorption systems for air separation.
PSA Simulation
System discretization Discretize space domain[1]
Finite volume method
Flux limiter
• Keep sharpness of steep fronts • Prevent oscillations
convert PDEs to ODEs
Smearing Oscillation 5
7 [3] Reza Haghpanah, et al. Multiobjective optimization of a 4-step adsorption process for post-combustion CO2 capture using nite volume technique.
WENO[3]
Differentiable
Two types of flux limiters
WENO performs similarly as van Leer to maintain accurate fronts
8
Without flux limiter, the simulation is poor near fronts.
No flux limiter VS van Leer WENO VS van Leer
Difference in CO2 concentration
Mole/m3
Difference in CO2 concentration
Mole/m3
WENO is computationally more efficient
Time to reach cyclic steady state
WENO Reduce 17.84% computational time
9
Optimization formulation[1]
Separation specific energy
Bed equations
Quality requirements
CO2 purity ≥92%
CO2 recovery ≥90%
Variable bounds
10
Periodic Boundary Condition Formulation[1]
• Add cyclic steady state equations
• Optimal variables: initial states and control variables
CCO2 CH2 qCO2 qH2 T
Step 1 Step 2 Step 3 Step 4 Step 5
Switch Beds and Repeat
Step 1
……
α β φ Pads Pdes ts Lbed
11
Convergence troubles[1]
• Sensitivity calculation -- Expensive (more than 95% computational time) -- Approximate within integrator tolerances, O(10-6 to 10-10)
• Hard to converge
-- Large number of iterations, computational time -- Infeasibility
• Sensitive to initial guess
MATLAB fmincon
SQP
Active-set
Interior-point
12 *First derivatives from direct sensitivity
020
4060
80100
120140
0
50
100
150
-2
-1
0
1
2
3
x 105
Notscaled Hessian
Optimization Diagnostics
Optimization ends up with no feasible solution
Final iteration 92
Objective value 85.08
Constraints violation 0.989
KKT violation 29.69
Large KKT violation is caused by φ
Check approximate Hessian matrix at final point, large elements also correspond to φ
13
Importance of scaling
Not scaled Scaled by 100 Scaled by 1000
Objective value 86.24 86.26 85.8
Constraints violation 0.289 9.78×10-4 2.32×10-3
KKT violation 29.42 3.338 8.598
Optimization data at iteration 80
Scale φ
With scaled φ, both primal and dual feasibility can be improved largely.
14
Redo optimization with all control variables except φ
Final iteration 232
Objective value 84.71
Constraints violation 7.68×10-7
KKT violation 4.39×10-3
Reduce problem dimension
Fix control variables optimization problem nonlinear equations solving problem # bed state variables = # bed equations
Optimization with only one type of control variables α β φ Pads Pdes ts Lbed
15
Conclusion & Future work
WENO flux limiter is more suitable for PSA optimization – high efficiency and better differentiability
Scaling improves optimizer performance
Simplified problems converge to optimum
o Continue optimization diagnostics
o Consider optimization with reduced order models
16