1. Overview & Applications
NGL recovery optimization maximizes plant profitability by balancing product recovery against operating costs while respecting equipment constraints and product specifications. The key decision is how much NGL to extract versus leaving in the residue gas stream.
Cryogenic plants
Turboexpander processes
GSP, SCORE, and similar processes achieve 85-98% ethane recovery using isentropic expansion.
Low-cost option
JT valve plants
Simple Joule-Thomson expansion achieves 20-50% C2 recovery with minimal capital.
Moderate recovery
Refrigeration systems
External propane or mixed refrigerant cooling provides 40-80% ethane recovery.
Maximum recovery
Enhanced processes
GSP with subcooling or recycle achieves 90-99% C2 recovery for rich gas streams.
NGL Plant Optimization Objective
Economic Objective Function:
Maximize: Net Margin = NGL Revenue - Operating Costs
NGL Revenue = Σ (Production_i × Price_i)
= (Q_C2 × P_C2) + (Q_C3 × P_C3) + (Q_C4+ × P_C4+)
Operating Costs = Cooling Cost + Compression Cost
= (Q_cooling × $/MMBtu) + (Power × $/kWh)
Subject to:
- Residue gas heating value ≥ 950 Btu/scf (pipeline spec)
- Residue gas hydrocarbon dewpoint ≤ cricondentherm
- Equipment capacity limits (columns, compressors, exchangers)
- Product specifications (vapor pressure, purity)
Key Optimization Variables
- Target C2 recovery: Primary decision variable - higher recovery = more revenue but higher costs
- Process temperatures: Cold separator, demethanizer top, reflux temperatures
- Operating pressures: Demethanizer pressure affects separation efficiency and recompression
- Feed split: Bypass fraction around turboexpander for temperature control
Economic significance: A 100 MMscfd plant with 5 GPM total NGL content produces ~12,000 bbl/day. At $40/bbl average NGL price, annual revenue is ~$175M. A 5% margin improvement from optimization yields $8-10M/year additional profit.
Process Selection Criteria
| Process Type | C2 Recovery | C3+ Recovery | Relative CAPEX | Best Application |
|---|---|---|---|---|
| JT Valve | 20-50% | 70-90% | Low (0.4×) | Low pressure inlet, small plants |
| External Refrigeration | 40-80% | 90-98% | Medium (0.7×) | Moderate recovery needs |
| Turboexpander (GSP) | 85-98% | 97-99.5% | High (1.0×) | High recovery, rich gas |
| Enhanced (GSP+) | 90-99% | 98-99.9% | Very High (1.3×) | Maximum recovery required |
IMAGE: NGL Recovery Process Comparison
Bar chart comparing 4 process types (JT Valve, Refrigeration, Turboexpander, Enhanced) with grouped bars for C2 Recovery %, C3+ Recovery %, and Relative CAPEX.
2. NGL Recovery Economics
Understanding NGL economics requires familiarity with GPM calculations, component physical properties, and the relationship between recovery level and operating costs. This section covers the fundamentals needed for economic optimization.
GPM - Gallons Per Mscf
GPM is the industry-standard measure of NGL content in natural gas. It represents the theoretical liquid volume recoverable per 1000 standard cubic feet of gas.
GPM Calculation (GPSA Section 16):
GPM = (mol% / 100) × (1000 scf/Mscf) × (1 lb-mol / 379.5 scf) × (gal / lb-mol)
Simplified:
GPM = mol% × (gal/lb-mol) / 37.95
Example - 8% Ethane:
GPM_C2 = 8 × 10.12 / 37.95 = 2.13 gal/Mscf
For 100 MMscfd feed:
Daily C2 potential = 2.13 × 100,000 = 213,000 gal/day = 5,071 bbl/day
IMAGE: GPM Calculation Diagram
Flowchart showing: Gas composition (mol%) → GPM per component → Total GPM → Production rate at various recoveries
NGL Component Physical Properties
Production calculations require accurate physical property data from GPA 2145 and GPSA Tables:
| Component | Formula | MW (lb/lb-mol) | Liquid Density (lb/gal) | gal/lb-mol | GPM per mol% |
|---|---|---|---|---|---|
| Ethane | C₂H₆ | 30.07 | 2.97 | 10.12 | 0.267 |
| Propane | C₃H₈ | 44.10 | 4.23 | 10.42 | 0.275 |
| i-Butane | i-C₄H₁₀ | 58.12 | 4.70 | 12.38 | 0.326 |
| n-Butane | n-C₄H₁₀ | 58.12 | 4.87 | 11.93 | 0.314 |
| Pentanes+ | C₅+ | 72.15 | 5.26 | 13.72 | 0.362 |
Shrinkage and Residue Gas
When NGL is extracted, the residue gas volume is reduced. This "shrinkage" affects pipeline nominations and gas sales contracts.
Shrinkage Calculation:
Shrinkage (%) = Σ (Component mol% × Recovery%)
Example:
Feed: 8% C2, 3% C3, 2% C4+
Recovery: 90% C2, 98% C3, 99% C4+
Shrinkage = (8 × 0.90) + (3 × 0.98) + (2 × 0.99)
= 7.2 + 2.94 + 1.98 = 12.12%
Residue Gas = 100 MMscfd × (1 - 0.1212) = 87.88 MMscfd
Ethane Rejection Economics
When ethane prices are low relative to operating costs, it may be more profitable to reject ethane into the residue gas stream. This "ethane rejection" mode is a key operational flexibility in modern NGL plants.
Ethane Rejection Decision:
Recover ethane if: C2 Revenue > Incremental Cost to Recover C2
Breakeven C2 Price = (Incremental Operating Cost) / (C2 Production Volume)
Example:
C2 production at 90% recovery: 5,000 bbl/day
Incremental cooling cost for C2 recovery: $15,000/day
Breakeven C2 price = $15,000 / (5,000 × 42 gal/bbl) = $0.071/gal
If market C2 price > $0.071/gal → Recover ethane
If market C2 price < $0.071/gal → Reject ethane
Typical breakeven range: $0.05 - $0.15/gal depending on plant efficiency
IMAGE: Ethane Rejection vs Recovery Economics
Chart showing: X-axis = Ethane price ($/gal), Y-axis = Net margin ($/day). Two lines: Full recovery mode vs Rejection mode, with crossover point at breakeven price.
Operational flexibility: Plants designed for ethane flexibility can switch between recovery and rejection modes within hours. This requires bypass piping around the demethanizer reflux system and control system modifications. The ability to reject ethane during low-price periods can improve annual margins by $2-5M for a 100 MMscfd plant.
Typical NGL Plant Economics
| Parameter | Lean Gas (2 GPM) | Moderate (4 GPM) | Rich Gas (6+ GPM) |
|---|---|---|---|
| NGL Production (bbl/MMscf) | ~50 | ~100 | ~150+ |
| Gross Revenue ($/MMscf) | $75-125 | $150-250 | $225-400 |
| Operating Cost ($/MMscf) | $25-50 | $40-80 | $50-100 |
| Net Margin ($/MMscf) | $50-75 | $100-170 | $175-300 |
| Preferred Process | JT or Refrigeration | Turboexpander | Enhanced Turboexpander |
3. Process Simulation Tools
Process simulators solve mass and energy balances, phase equilibrium, and equipment performance equations to predict plant behavior under different operating conditions.
Industry-Standard Simulators
| Simulator | Vendor | Primary Use | Optimization Capabilities |
|---|---|---|---|
| Aspen HYSYS | AspenTech | Gas processing, LNG, upstream | Built-in optimizer, case studies, sensitivity analysis |
| Aspen Plus | AspenTech | Refining, petrochemicals, chemicals | Optimization solver, design spec, calculator blocks |
| PRO/II | Schneider (AVEVA) | Refining, gas processing | Optimizer, sensitivity analysis |
| ProMax | Bryan Research | Gas processing, pipelines, NGL | Case studies, what-if analysis |
| UniSim Design | Honeywell | Oil & gas, refining | Optimizer, spreadsheet interface |
| CHEMCAD | Chemstations | Chemical plants, gas processing | Optimization add-on |
Simulation-Based Optimization Workflow
Optimizer-Simulator Interface:
1. Optimizer proposes new decision variable values x_trial
2. Simulator runs with x_trial to compute outputs y
3. Evaluate objective function f(x) and constraints g(x), h(x)
4. Optimizer updates x based on gradient or search method
5. Repeat until convergence (optimal solution found)
Convergence criteria:
|f(x_k) - f(x_k-1)| < tolerance (objective improvement small)
||∇f(x)|| < tolerance (gradient near zero)
All constraints satisfied
Built-In Optimization Tools
Aspen HYSYS Optimizer
- Solver: Sequential Quadratic Programming (SQP) for NLP problems
- Variables: Any stream property, equipment parameter, or flowsheet variable
- Objective: Maximize/minimize any calculated variable (profit, energy, emissions)
- Constraints: Linear/nonlinear inequality and equality constraints
- Typical use: Maximize NGL recovery in gas plant, minimize reboiler duty in distillation
Aspen Plus Optimization
- Design Spec: Single-variable optimization (vary one parameter to meet one target)
- Optimization Block: Multi-variable optimization with constraints
- Sensitivity Analysis: Systematic variation of inputs, plot outputs
- Case Studies: Tabulate results for multiple scenarios
Example: Gas Plant Optimization
Optimize a cryogenic gas plant to maximize NGL recovery while meeting pipeline gas spec:
Decision Variables (x):
- Demethanizer top temperature (°F)
- Demethanizer pressure (psia)
- JT valve upstream pressure (psia)
- Cold separator temperature (°F)
Objective Function (maximize):
f(x) = NGL revenue - fuel gas cost - compression power cost
NGL revenue = Q_NGL × (P_ethane × y_C2 + P_propane × y_C3 + ...)
Fuel cost = Q_fuel × P_fuel
Power cost = W_comp × P_electric / η_motor
Constraints:
Residue gas heating value ≥ 950 Btu/scf (pipeline spec)
Residue gas hydrocarbon dewpoint ≤ -20°F at 800 psia
Demethanizer bottom temp > hydrate formation temp + 10°F
Compressor discharge pressure ≤ 1200 psia (equipment limit)
Reboiler duty ≤ 20 MMBtu/hr (equipment limit)
NGL vapor pressure ≤ 200 psia at 100°F (storage spec)
Typical Result:
Base case: 85% C2 recovery, 98% C3+ recovery
Optimized: 90% C2 recovery, 99% C3+ recovery
Incremental profit: $2-5 million/year for 100 MMscfd plant
Surrogate Models for Fast Optimization
Rigorous simulations are computationally expensive (minutes per run). For real-time optimization or Monte Carlo analysis, use surrogate models:
- Polynomial regression: f(x) = a₀ + a₁x₁ + a₂x₂ + a₁₂x₁x₂ + a₁₁x₁² + ... (fast, limited accuracy)
- Neural networks: Multi-layer perceptron trained on simulation data (high accuracy, black box)
- Kriging (Gaussian process): Interpolates between simulation runs, provides uncertainty estimate
- Reduced-order models: Simplified physics-based model capturing key phenomena
Model fidelity tradeoff: Rigorous simulators are accurate but slow (1-10 min/run). Surrogate models are fast (milliseconds) but require training data and may be less accurate outside training range. Use rigorous models for design, surrogates for real-time optimization.
4. Constraint Analysis
Constraint analysis identifies bottlenecks limiting plant performance. Removing or relaxing the active constraint provides the largest incremental improvement.
Active vs Inactive Constraints
Constraint Status:
Active constraint: g_i(x*) = 0 (constraint is binding at optimal solution)
Inactive constraint: g_i(x*) < 0 (constraint not limiting)
At the optimal solution, at least one constraint is typically active.
Example:
Maximize throughput in gas plant
Constraints:
1. Inlet compressor power ≤ 5000 HP
2. Demethanizer reboiler duty ≤ 20 MMBtu/hr
3. Residue gas heating value ≥ 950 Btu/scf
If optimal throughput is limited by reboiler duty = 20 MMBtu/hr,
then reboiler constraint is ACTIVE (binding).
Inlet compressor uses 4200 HP → INACTIVE (slack = 800 HP).
Relaxing active constraint (increase reboiler capacity) allows higher throughput.
Relaxing inactive constraint (add compressor HP) has no immediate benefit.
Lagrange Multipliers and Shadow Prices
The Lagrange multiplier (shadow price) quantifies the incremental benefit of relaxing a constraint:
Lagrangian Function:
L(x, λ, μ) = f(x) + Σ λ_i g_i(x) + Σ μ_j h_j(x)
Where:
λ_i = Lagrange multiplier for inequality constraint i
μ_j = Lagrange multiplier for equality constraint j
At optimal solution:
∇_x L = 0 (gradient with respect to decision variables)
Shadow price = λ_i = ∂f*/∂b_i
If constraint is g_i(x) ≤ b_i, then:
λ_i = incremental improvement in objective per unit increase in b_i
Example:
Reboiler duty constraint: Q_reb ≤ 20 MMBtu/hr
Shadow price: λ = $500/yr per Btu/hr additional capacity
→ Increasing reboiler capacity by 1 MMBtu/hr gains $500,000/yr profit
Use shadow prices to prioritize capital projects (debottlenecking).
Common Process Constraints
| Constraint Type | Example | Debottlenecking Options |
|---|---|---|
| Equipment capacity | Compressor power, column diameter, heat exchanger area | Upgrade equipment, add parallel unit, increase efficiency |
| Utility limits | Cooling water, steam, fuel gas, electric power | Add utility capacity, improve heat integration, use waste heat |
| Product specifications | RVP, sulfur, heating value, freeze point | Adjust operating conditions, add treating capacity, blend optimization |
| Environmental permits | Emissions (NOx, VOC, CO2), flaring, wastewater discharge | Modify permit, add abatement equipment, operational changes |
| Safety systems | Relief valve capacity, flare capacity, trip setpoints | Upgrade relief system, improve process control, reduce inventory |
| Feedstock availability | Gas supply volume, crude slate, catalyst inventory | Secure additional supply, flexible feed design, contracts |
Constraint Propagation
Constraints interact through material and energy balances. Changing one constraint may shift the bottleneck elsewhere:
- Sequential bottlenecks: After removing constraint A, constraint B becomes active
- Coupled constraints: Two constraints interact (e.g., column diameter and reboiler duty both limit throughput)
- Redundant constraints: Multiple constraints limit the same variable (most restrictive one is active)
Sensitivity Analysis
Systematically vary parameters to understand their impact on the objective:
One-at-a-Time Sensitivity:
Fix all variables except one, vary that variable over range:
x_i = x_i,base ± Δx_i
Plot f(x) vs x_i to visualize sensitivity
Steep slope → high sensitivity → priority optimization variable
Flat slope → low sensitivity → insensitive parameter
Tornado Diagram:
Rank parameters by impact on objective:
1. Feed composition: ±$2M/yr
2. Ethane price: ±$1.5M/yr
3. Fuel gas cost: ±$0.8M/yr
4. Compressor efficiency: ±$0.5M/yr
Focus optimization efforts on high-impact variables.
Constraint analysis strategy: Identify the active constraint limiting performance. Calculate shadow price to quantify incremental benefit of relaxing that constraint. Prioritize capital projects and operational changes based on shadow prices and project costs.
5. Linear & Nonlinear Programming
Mathematical programming techniques solve structured optimization problems efficiently using gradient-based or search algorithms.
Linear Programming (LP)
LP Standard Form:
Maximize: c^T x
Subject to:
A x ≤ b (linear inequality constraints)
x ≥ 0 (non-negativity)
Where:
x = decision variables (n-dimensional vector)
c = objective coefficients (revenue/cost per unit)
A = constraint matrix (m × n)
b = constraint right-hand sides
Example: Blending Optimization
Blend 3 crude oils to maximize profit:
Decision variables:
x₁ = volume of crude A (bbl/day)
x₂ = volume of crude B (bbl/day)
x₃ = volume of crude C (bbl/day)
Objective (maximize profit):
f = 50x₁ + 45x₂ + 55x₃ ($/day)
Constraints:
x₁ + x₂ + x₃ ≤ 100,000 (total capacity, bbl/day)
0.25x₁ + 0.30x₂ + 0.20x₃ ≥ 25,000 (gasoline yield, bbl/day)
x₁ ≤ 40,000 (crude A availability)
x₂ ≤ 50,000 (crude B availability)
x₃ ≤ 60,000 (crude C availability)
x₁, x₂, x₃ ≥ 0
LP solution methods: Simplex algorithm, interior-point methods
Commercial solvers: CPLEX, Gurobi, GLPK
Nonlinear Programming (NLP)
When objective or constraints contain nonlinear terms (products, powers, logarithms, exponentials), use NLP methods:
NLP Problem:
Minimize: f(x) (nonlinear objective)
Subject to:
g_i(x) ≤ 0 i = 1...m (nonlinear inequalities)
h_j(x) = 0 j = 1...n (nonlinear equalities)
Example: Distillation Column Optimization
Minimize total annual cost:
f(x) = Capital cost + Operating cost
= K × D^0.65 × N^0.5 + C_steam Q_reb + C_cooling Q_cond
Decision variables:
D = column diameter (ft)
N = number of trays
R = reflux ratio
P = column pressure (psia)
Constraints (nonlinear):
Product purity ≥ 95% (depends on N, R, P via Fenske-Underwood equations)
Flooding velocity < 80% (depends on D, vapor rate, liquid rate)
Temperature profile feasible (bubble point < tray temp < dew point)
NLP solvers: SQP (Sequential Quadratic Programming), interior-point, GRG (Generalized Reduced Gradient)
Common NLP Solution Methods
| Method | Algorithm | Pros | Cons |
|---|---|---|---|
| SQP (Sequential Quadratic Programming) | Solve quadratic approximation of Lagrangian iteratively | Fast convergence, handles constraints well | Requires gradient calculation, may find local optimum |
| Interior-Point | Follow central path to optimal solution | Efficient for large problems, robust | Requires good starting point |
| GRG (Generalized Reduced Gradient) | Reduce problem to unconstrained subproblem | Handles equality constraints naturally | Slower than SQP for large problems |
| Genetic Algorithm | Evolutionary search, population-based | Global optimization, derivative-free | Slow, many function evaluations |
| Particle Swarm Optimization | Swarm intelligence, collaborative search | Simple to implement, derivative-free | No convergence guarantee, tuning required |
Gradient Calculation Methods
Gradient-based NLP solvers require derivatives ∂f/∂x and ∂g/∂x:
- Analytical derivatives: Hand-derive equations (accurate, fast evaluation, tedious to implement)
- Finite differences: Approximate ∂f/∂x ≈ [f(x+ε) - f(x)] / ε (easy to implement, inaccurate for small ε, 2n function evaluations)
- Complex-step method: ∂f/∂x ≈ Im[f(x + iε)] / ε (machine-precision accuracy, requires complex arithmetic)
- Automatic differentiation: Compiler computes exact derivatives (accurate, fast, requires AD-enabled software)
Global vs Local Optimization
Local Optimum:
f(x*) ≤ f(x) for all x in neighborhood of x*
Gradient-based methods converge to local optimum
Global Optimum:
f(x*) ≤ f(x) for all feasible x
No guarantee that local optimum is global optimum for nonconvex NLP
Strategies for global optimization:
1. Multi-start: Run optimizer from multiple random initial guesses, select best result
2. Simulated annealing: Probabilistic method accepting worse solutions to escape local optima
3. Branch-and-bound: Systematically partition search space, prune infeasible regions
4. Convex relaxation: Replace nonconvex problem with convex approximation, solve to get bounds
For process optimization, multi-start SQP is common practice:
- Run 10-50 optimizations from random initial points
- Select solution with best objective value
- Verify constraints are satisfied and solution is physically realistic
Mixed-Integer Programming (MILP/MINLP)
Some decisions are discrete (on/off, equipment selection, routing):
- Binary variables: y ∈ {0, 1} (unit on/off, use technology A or B)
- Integer variables: n ∈ {1, 2, 3, ...} (number of trays, number of parallel units)
- Applications: Plant design, scheduling, supply chain optimization, retrofit decisions
- Solution methods: Branch-and-bound, branch-and-cut, outer approximation (for MINLP)
Method selection: Use LP for blending, scheduling, and allocation problems (fast, global optimum). Use NLP for equipment sizing and process operating conditions (nonlinear physics). Use MINLP for design problems with discrete decisions (computationally expensive, use for capital projects).
6. Real-Time Optimization (RTO)
Real-Time Optimization continuously monitors plant performance and adjusts setpoints to maintain optimal operation as feed composition, product prices, and equipment performance change.
RTO System Architecture
RTO Workflow:
1. Data Reconciliation:
- Collect measurements from DCS/SCADA (flows, temperatures, pressures, compositions)
- Reconcile measurements to satisfy mass/energy balances (remove sensor errors)
- Gross error detection (identify faulty instruments)
2. Parameter Estimation:
- Update model parameters to match current plant performance
- Examples: Heat transfer coefficients, tray efficiencies, reaction kinetics
- Minimize difference between model predictions and measurements
3. Optimization:
- Solve NLP to maximize profit subject to constraints
- Decision variables: setpoints for temperatures, pressures, flow rates
- Constraints: Equipment limits, product specs, safety margins
4. Implementation:
- Send optimal setpoints to regulatory control layer (DCS)
- Ramp setpoints gradually (rate limits) to avoid upsetting process
- Monitor for constraint violations or model mismatch
5. Performance Monitoring:
- Track economic benefit vs baseline operation
- Detect model degradation (need to re-tune parameters)
- Alert operators if optimization is infeasible
Typical RTO cycle time: 15-60 minutes
Data Reconciliation
Process measurements contain random errors and bias. Data reconciliation adjusts measurements to satisfy conservation laws:
Data Reconciliation Problem:
Minimize: Σ w_i (x_i - x_i,measured)² (weighted least squares)
Subject to:
A x = 0 (mass balance constraints)
B x = 0 (energy balance constraints)
Where:
x = reconciled (adjusted) measurements
x_measured = raw measurements from sensors
w_i = weight (inverse variance) of sensor i
A, B = balance matrices
Example:
Measured flow rates: F_in = 100 ± 2%, F_out1 = 55 ± 3%, F_out2 = 42 ± 3%
Mass balance: F_in = F_out1 + F_out2
100 ≠ 55 + 42 (imbalance of 3 units)
Data reconciliation adjusts values to:
F_in = 98.5, F_out1 = 54.5, F_out2 = 44.0
Now 98.5 = 54.5 + 44.0 ✓
Use reconciled values for optimization (more accurate than raw measurements).
Model-Plant Mismatch
Process models are never perfect. Manage mismatch through:
- Bias correction: Add offset to model predictions to match current plant data (simple, temporary fix)
- Parameter adaptation: Re-tune model parameters (heat transfer coefficients, efficiencies) to match plant performance
- Constraint back-off: Add safety margin to constraints to account for model uncertainty (e.g., operate at 95% of column flooding instead of 100%)
- Model re-identification: Periodically rebuild model using recent plant data (months to years)
RTO Implementation Challenges
| Challenge | Impact | Mitigation |
|---|---|---|
| Sensor failures | Bad data → infeasible optimization | Gross error detection, sensor redundancy, virtual sensors |
| Model inaccuracy | Optimal setpoints are suboptimal in reality | Parameter estimation, bias updates, constraint back-off |
| Disturbances | Plant drifts from optimal setpoint between RTO cycles | MPC (Model Predictive Control) for fast regulatory control |
| Infeasible solutions | No feasible setpoint satisfying all constraints | Constraint relaxation (penalty functions), alert operator |
| Slow convergence | RTO takes too long, plant conditions change during optimization | Use surrogate model, warm-start solver, faster hardware |
| Operator trust | Operators override RTO recommendations | Operator training, demonstrate economic benefit, gradual rollout |
Economic Performance Monitoring
Quantify RTO benefit by comparing actual operation to baseline (pre-RTO) or steady-state optimization:
RTO Economic Benefit:
Benefit = Actual profit - Baseline profit
Or:
Benefit = Optimal profit (steady-state) - Opportunity loss
Where:
Opportunity loss = Σ (shadow price × constraint violation)
Example:
RTO optimizes ethane recovery in gas plant
Baseline: 85% C2 recovery, $10M/year profit
RTO average: 90% C2 recovery, $12M/year profit
Benefit: $2M/year
KPIs to track:
- Average % of time RTO is active (target > 90%)
- Average objective function value (profit, throughput, energy)
- Frequency of constraint violations (target < 5% of time)
- Model-plant mismatch (residuals between predicted and measured values)
Integration with Advanced Process Control (APC)
RTO provides economic setpoints; APC (Model Predictive Control) regulates process to track those setpoints:
- RTO layer: Slow (15-60 min cycle), economic optimization, steady-state model
- APC/MPC layer: Fast (1-5 min cycle), regulatory control, dynamic model, constraint handling
- Regulatory control (PID): Very fast (seconds), single-loop controllers, implement APC outputs
RTO best practices: Start with rigorous offline optimization to establish baseline. Implement RTO in phases (data reconciliation first, then parameter estimation, then closed-loop optimization). Monitor performance continuously and re-tune model as needed. Typical RTO payback: 6-18 months for large facilities.
RTO Case Study: NGL Fractionation Train
Facility: 100,000 bbl/day NGL fractionation (deethanizer, depropanizer, debutanizer)
RTO Objectives:
Maximize: Total margin = Σ(Product flow × Price) - Utility costs - Feed cost
Decision Variables (18 total):
- Column pressures (3)
- Reflux ratios (3)
- Reboiler duties (3)
- Feed split ratios (2)
- Side draw rates (7)
Constraints (42 total):
- Product specifications (ethane C3 < 5%, propane C2 < 2%, C4 < 2.5%)
- Column flooding (vapor velocity < 80% of flood)
- Reboiler/condenser duties within equipment limits
- Compressor power < 8000 HP
- Minimum approach temperature in exchangers
Results:
Baseline (manual operation): $45M/year margin
RTO optimal (average): $48M/year margin
Benefit: $3M/year (6.7% improvement)
RTO system cost: $2M (hardware, software, engineering)
Payback: 8 months
Key insight: RTO adjusted column pressures to balance between separation efficiency (lower P → more trays → better separation) and compression cost (lower P → higher recompression power). Operators had been running columns at constant pressure.
Ready to use the calculator?
→ Launch Calculator