Chemical / Process Operations

Distillation Column Parameter Estimation

Dynamic parameter estimation and soft sensing for separation process monitoring

Introduction

Distillation columns are ubiquitous in chemical, petrochemical, and pharmaceutical industries for separating liquid mixtures. Control and optimization require accurate models, but key parameters—tray efficiencies, heat transfer coefficients, and feed composition—drift due to fouling, corrosion, and upstream variability.

The challenge: process control systems rely on infrequent lab composition measurements (gas chromatography every 2–8 hours), while temperature sensors provide continuous data. Can we estimate hidden composition profiles and uncertain process parameters from temperature measurements alone?

      Why this matters: Accurate soft sensors reduce lab analysis burden by 50–70%, enable real-time optimization of reflux ratio for energy savings, and provide early warnings of tray damage or feed upsets—preventing off-spec product and unplanned shutdowns.
    

Process Science: Vapor-Liquid Equilibrium and Mass Balance

A distillation column separates a binary mixture (e.g., methanol-water, benzene-toluene) by exploiting differences in volatility. Key physics:

Vapor-liquid equilibrium (VLE): At each tray, liquid and vapor reach thermodynamic equilibrium described by Raoult's law or more complex activity coefficient models
Mass transfer: Component transfer between phases governed by tray efficiency (Murphree efficiency EMV)
Energy balance: Condenser duty, reboiler duty, and sensible/latent heat determine vapor flow rates
Hydraulic dynamics: Liquid holdup on each tray creates time delays; vapor flows are quasi-steady-state
Feed disturbances: Upstream process variability changes feed flow rate F and composition zF

Governing Equations: Stagewise DAE Model

We use a MESH (Material, Equilibrium, Summation, Heat) formulation for N trays:

States (per tray $j$): $$\begin{aligned} x_j &= \text{liquid mole fraction of light component [-]} \\ M_j &= \text{liquid holdup [kmol]} \\ T_j &= \text{tray temperature [K]} \end{aligned}$$ Differential Equations (Mass): $$\begin{aligned} \frac{dM_j}{dt} &= L_{j+1} + F \cdot \delta(j-j_F) - L_j - V_j \quad \text{[total moles]} \\ \frac{d(M_j \cdot x_j)}{dt} &= L_{j+1} \cdot x_{j+1} + F \cdot z_F \cdot \delta(j-j_F) - L_j \cdot x_j - V_j \cdot y_j \quad \text{[light component]} \end{aligned}$$ Algebraic Equations: $$\begin{aligned} y_j &= E_{\text{MV}} \cdot y_j^* + (1 - E_{\text{MV}}) \cdot y_{j+1} \quad \text{[Murphree efficiency]} \\ y_j^* &= K_j(T_j, P) \cdot x_j \quad \text{[VLE equilibrium, } K = \text{relative volatility]} \\ L_j &= \beta \cdot \sqrt{M_j - M_w} \quad \text{[Francis weir formula, hydraulic]} \\ V_j &= V_{\text{total}} \quad \text{[constant molar overflow assumption]} \end{aligned}$$ Energy Balance (simplified): $$\begin{aligned} 0 &= L_{j+1} \cdot h_L(T_{j+1}) + V_{j-1} \cdot h_V(T_{j-1}) - L_j \cdot h_L(T_j) - V_j \cdot h_V(T_j) + F \cdot h_F \cdot \delta(j-j_F) \\ &\Rightarrow T_j = g(x_j, P, V_j, L_j) \quad \text{[temperature from enthalpy balance]} \end{aligned}$$ Boundary Conditions: $$\begin{aligned} j=1 \text{ (condenser)}: & \quad L_1 = RD \text{ (reflux)}, \; V_1 = 0 \\ j=N \text{ (reboiler)}: & \quad V_N = V_{\text{total}}, \; L_{N+1} = B \text{ (bottoms flow)} \end{aligned}$$

Unknown Parameters (8 total)

Parameter	Description	Units	Typical Range
EMV	Murphree vapor tray efficiency	—	0.5 – 0.95
α	Relative volatility (temperature-dependent)	—	1.2 – 4.0
Mw	Weir holdup (hydraulic)	kmol	0.1 – 2.0
β	Weir coefficient	—	0.5 – 1.5
UA_cond	Condenser heat transfer coefficient	kW/K	5 – 50
UA_reb	Reboiler heat transfer coefficient	kW/K	5 – 50
zF	Feed composition (drifts with upstream)	—	0.3 – 0.7
F	Feed flow rate (measured but biased)	kmol/h	50 – 200

Parameter Estimation Problem

Given measurements from a distillation column:

Continuous: Tray temperatures T1, T5, T10, T15, T20 [thermocouples, 1-min resolution]
Continuous: Reflux rate RD, reboiler duty QR [flow meters, heat flux sensors]
Delayed: Distillate composition xD, bottoms composition xB [GC every 4 hours]

The inverse problem is formulated as moving horizon estimation (MHE) with a sliding window of τ = 8 hours:

Minimize over window $[t-\tau, t]$: $$J(\theta, x_0) = ||x(t-\tau) - \hat{x}_0||^2_P + \sum ||y_{\text{obs}}(t_k) - h(x(t_k), \theta)||^2_R + \lambda \cdot ||\theta - \theta_{\text{prior}}||^2$$ Subject to: $$\begin{aligned} \frac{dx}{dt} &= f(x, u, \theta) \quad \text{[DAE model integration]} \\ x_j &\in [0, 1], \; M_j > 0, \; T_j \in [T_{bp1}, T_{bp2}] \quad \text{[physical bounds]} \end{aligned}$$ Where: $$\begin{aligned} \theta &= [E_{\text{MV}}, \alpha, M_w, \beta, UA_{\text{cond}}, UA_{\text{reb}}, z_F, F] \\ y_{\text{obs}} &= [T_1, T_5, T_{10}, T_{15}, T_{20}, x_D^*, x_B^*] \quad \text{[* = sparse]} \\ h(x, \theta) &= \text{measurement function (temperatures from energy balance)} \\ x_0 &= \text{initial state estimate at } t-\tau \\ P, R &= \text{prior covariance, measurement noise covariance} \end{aligned}$$

      Soft sensor output: The MHE framework simultaneously estimates unmeasured tray compositions x3, x7, x12, ... and uncertain parameters θ, providing real-time composition profiles without waiting for lab results.
    

Simulation Framework

Generate synthetic column operation data:

Nominal steady-state: Solve algebraic MESH equations for initial profiles (xj, Mj, Tj)
Step disturbances:
- Feed composition: zF steps from 0.45 to 0.52 at t=2h
- Feed flow rate: F steps from 100 to 120 kmol/h at t=5h
- Reflux ratio: RD/D increases 10% at t=7h (energy optimization)
DAE integration: Use DASSL or IDA solver with consistent initial conditions
Add noise: Temperature: ±0.3 K, GC composition: ±0.01 mole fraction, delayed by 4h
Parameter drift: EMV decreases 5% over 24h (tray fouling), zF oscillates ±0.03 (upstream variability)

Estimation Workflow

1. Dynamic Least Squares (Offline Calibration)

For historical data campaigns:

Batch optimization over entire run (24–72 hours)
Simultaneous parameter estimation and trajectory smoothing
Use SQP or interior-point methods with DAE constraints

2. Moving Horizon Estimation (MHE)

For real-time soft sensing:

At each time step $t$: $$\begin{aligned} &\text{1. Slide window: discard data before } t-\tau \\ &\text{2. Solve constrained NLP:} \\ &\qquad \min J(\theta, x(t-\tau)) \text{ over } [t-\tau, t] \\ &\text{3. Extract current estimates: } \hat{\theta}(t), \; \hat{x}(t) \\ &\text{4. Advance window to } [t-\tau+\Delta t, t+\Delta t] \\[1em] &\text{Computational budget: }\sim 10 \text{ seconds per update (DAE solve + NLP)} \\ &\qquad \Rightarrow \text{ Use warm-start from previous solution} \end{aligned}$$

3. Extended Kalman Filter (EKF) for State Estimation

Faster alternative when parameters are known or slowly varying:

Augment state vector: [x1, ..., xN, M1, ..., MN, EMV, zF]
Linearize DAE system around current estimate
Kalman update when delayed GC measurements arrive
Computational cost: ~0.5 seconds per update (suitable for 1-min control loops)

Identifiability

Key insights from sensitivity analysis:

      Temperature-only limitations:
      Tray efficiency EMV vs. relative volatility α: Highly correlated—both affect separation sharpness. Requires occasional GC measurements to distinguish.
Feed composition zF: Observable from temperature profile shift, but 2–4 hour delay until column responds (tray holdups create time constants)
Hydraulic parameters (Mw, β): Weakly observable from steady-state temperatures. Require transient excitation (step changes in reflux).
Heat transfer coefficients: Identifiable only if condenser/reboiler temperatures are measured separately

    

Recommendation: Fix hydraulic parameters (Mw, β) from startup commissioning tests. Estimate EMV, zF online using MHE with weekly GC recalibration. Monitor UA_cond/UA_reb for fouling trends (annual maintenance indicator).

Validation Strategy

Validation Test	Method	Acceptance Criterion
Temperature prediction	Hold-out test: predict T5, T15 from T1, T10, T20	RMSE < 0.5 K
Soft sensor accuracy	Compare MHE composition estimate to delayed GC	\|xD_est - xD_GC\| < 0.015
Disturbance rejection	Feed step change: time to detect zF change	Detection within 30 minutes
Robustness	Operate with 1 failed thermocouple	Graceful degradation, no divergence
Computational speed	MHE solve time on PLC hardware	< 10 seconds per update

Industrial Value

Mechanistic parameter estimation and soft sensing deliver:

Energy savings: Real-time optimization of reflux ratio based on accurate composition estimates reduces steam consumption by 8–15% (typical payback: 6–18 months)
Product quality: Early detection of off-spec distillate/bottoms enables corrective action before product tank contamination
Maintenance planning: Monitoring tray efficiency EMV trends identifies fouling or damage 2–3 weeks before product quality deteriorates
Lab efficiency: Reducing GC frequency from every 2h to every 8h cuts analytical costs by 75% while maintaining control performance
Startup acceleration: Model-based composition estimates reduce time to steady-state by 20–30% (saving 4–6 hours of transition per startup)

      Case study: A petrochemical refinery implemented MHE-based soft sensors across 12 distillation columns. Energy savings totaled $2.4M annually, and unplanned shutdowns due to quality excursions decreased by 60% (avoiding $5M in lost production).
    

Try ProcessLM

ProcessLM automates the entire workflow: describe your column configuration and control objectives in natural language, and it generates the MESH model, tunes the MHE problem, and deploys soft sensors—no Aspen Custom Modeler scripting required.

Request Early Access