Real-Time Optimisation That Fits: An NGL Extraction Case With Surrogate Models and Modifier Adaptation
- 5 days ago
- 24 min read
Real-Time Optimisation (RTO) is frequently discussed in industry, but not so often presented in a way that links theory, implementation and live operation. This work provides an overview of this industrial technology, clarifies its role within the process control hierarchy, and demonstrates live operation of a specific RTO class based on Modifier Adaptation (MA-RTO).
The implemented RTO relies on machine learning surrogate models to optimise simulated NGL Extraction process in LNG scrub columns. Its distinctive feature is real-time adaptation to changing plant performance. As demonstrated in the work, such adaptation enables convergence to the true plant optimum even in the presence of model-plant mismatch.
The demo is configured on a web-hosted instance of ARTData where engineers and students can observe, experiment, and learn how complex control logic works in real time environment. ARTData provides the execution layer, communication interface, and visualisation tools, making it possible to run the RTO in a framework that resembles industrial deployment. This implementation framework can be considered a lightweight and cost-effective deployment option for smaller-scale optimisation problems, where a full enterprise-grade RTO infrastructure may not be economically justified.
The post consists of five sections covering the following aspects:
What Makes an RTO – theoretical background and positioning within the control hierarchy;
The Demo Process: NGL Extraction in an LNG Scrub Column – description of the optimised process and its economics;
ARTData RTO Architecture and Implementation – technical details including the MA-RTO loop structure;
ARTData RTO Operation and Results – comparison of optimisation outcomes for “known” and “unknown” plant behaviour;
Key Takeaways – practical lessons and implementation insights from the project.
To get the most out of this post, it’s best to follow along using the live ARTData demo environment and Quick Reference Guide accessible via below links:
1. What Makes an RTO
1.1 Where RTO Sits in the Control Hierarchy
Real plants in the process industries rarely operate exactly at their design operating points. One reason is the continuous adjustment of operating conditions in response to varying feed quality, required throughput, product specifications, ambient conditions, and other disturbances. Another reason is the inevitable mismatch between the real plant and the model used during the design stage. In practice, plants always deviate from design to some degree.
How, then, can we ensure that a plant continues to operate optimally when there is no longer a reliable design reference? Real-Time Optimisation (RTO) emerged in the late 1980s and early 1990s as a process control methodology to address this question. Its purpose is to maximise the economic performance of operating plants on a repetitive basis, despite limited model accuracy and plant-model mismatch [1] [2].
RTO solves a steady-state optimisation problem: it determines the most economically efficient operating conditions and passes them as setpoints to a lower control layer responsible for dynamic control, for example Model-Predictive Control (MPC) or base-layer regulatory control. The optimisation is performed online using a steady-state process model, current plant measurements, economic information, and explicit operating constraints.
Within the process control hierarchy, RTO typically sits at Level 3, between offline planning and scheduling activities above, and supervisory or advanced control below (Fig. 1.1).
![Fig. 1.1 - Levels of process control and optimisation as per ISA-95 functional model [3].](https://static.wixstatic.com/media/b2efc2_0210367aae544abca4d5868cdefaa54f~mv2.png/v1/fill/w_980,h_521,al_c,q_90,usm_0.66_1.00_0.01,enc_avif,quality_auto/b2efc2_0210367aae544abca4d5868cdefaa54f~mv2.png)
While regulatory control operates on a seconds time scale, RTO runs at a higher control layer with execution intervals of minutes or hours. RTO requires information about operating constraints to ensure that its solution remains feasible and within the operating window. However, constraint enforcement and other dynamic control tasks are handled at the regulatory control level. RTO does not replace it. Therefore, proper alignment of constraints between the layers is essential.
1.2 Differentiation of RTO from MPC and Offline Optimisation
MPC, the most widely used supervisory control technique, determines and drives the plant along optimal dynamic trajectories using simplified process models, typically linear ones. Its primary objective is to track setpoints of controlled variables while keeping them within specified limits. MPC can handle large multivariable systems and natively incorporates operational constraints into the control problem [4].
If an economic objective function can be expressed directly in terms of MPC variables, it may be optimised together with the dynamic control problem. In cases where the economically optimal operating point lies on active constraints, and the linear model remains sufficiently accurate in that region, RTO may not provide additional value. In many practical situations, however, the optimum does not lie exactly on constraints, or the economic objective cannot be represented accurately using a simple linear model. In such cases, RTO closes this gap by leveraging a more accurate nonlinear steady-state model that is repeatedly calibrated using plant measurements at steady-state conditions.
Continuous process model adaptation is a fundamental component of RTO and distinguishes it from offline optimisation or online advanced setpoint scheduling. Online adaptation ensures that the gradients of the objective function (partial derivatives with respect to manipulated variables) remain as accurate as possible, which is a necessary condition for optimality [5]. Plant measurements collected at steady state are used to update and align the model, which is then used for economic optimisation. Once the plant reaches the new optimal operating conditions and stabilises, the cycle is repeated. In this way, RTO forms a closed loop with the plant, whereas offline optimisation and setpoint scheduling operate in an open-loop manner without feedback, and therefore without guaranteed convergence to the true economic optimum.
1.3 Classical Two-Step MPA-RTO Algorithm
There are several RTO methodologies, differing primarily in how plant-model mismatch is addressed. Plant-model mismatch arises from three main types of uncertainty [6]:
Parametric uncertainty – when model parameters do not accurately reflect plant reality but can be recalculated from plant measurements to recalibrate the model. The heat transfer coefficient of a heat exchanger is a typical example.
Structural uncertainty – when the model omits an important aspect of the real process due to incorrect assumptions or unknown phenomena, for example changes in reaction kinetics.
Process disturbances and measurement uncertainty – when the plant is undergoing a dynamic transition or when measurements contain noise and errors.
The classical two-step RTO approach, also known as Model Parameter Adaptation (MPA), addresses only parametric uncertainty. It relies on plant measurements to update selected model parameters in the first step, after which the updated model is used for optimisation in the second step [7] [8]. The corresponding block diagram is shown on Fig. 1.2.
At the beginning of each cycle, the RTO application retrieves historical measurement data from the PCS over a specified lookback window. The collected data may require pre-processing such as scaling or unit conversion. Abnormal or invalid data is filtered out at this stage.
It is essential that the data used in subsequent steps corresponds to steady-state operating conditions. Several steady-state detection techniques, typically based on statistical analysis, are used to evaluate data variation and trends over time and determine whether the plant is sufficiently stable.
Where applicable, mass or energy balances are enforced on the measured data through a reconciliation procedure. The most reliable measurements and fixed streams (e.g. flaring) are treated as hard constraints, and any calculated imbalance is distributed among the remaining streams. This ensures that parameter calibration is based on reconciled and physically consistent data.
In the next step of MPA-RTO, key model parameters are updated to match current plant behaviour. These typically include heat transfer coefficients, catalyst activity, fouling factors, furnace efficiency etc. Parameter updating compensates for equipment degradation and gradual process changes. Once parameter estimation is complete, the model is updated and optimal operating conditions are determined by solving an economic optimisation problem.
The MPA-RTO approach does not address structural uncertainty and therefore does not guarantee that the model-based optimum will be optimal for the plant. Because achieving and verifying model adequacy is challenging, this method often motivates the development of increasingly detailed models in the hope of improving accuracy. In practice, this can lead to significant complexity, difficulty in obtaining the necessary process information, and substantial engineering effort to maintain the model. Without continuous maintenance, plant-model mismatch can quickly accumulate.
1.4 Key Challenges
Several challenges complicate the successful deployment and long-term maintenance of RTO applications. Some are specific to particular RTO methodologies, while others are common to all approaches [9] [10].
Challenge 1. Mismatch between the RTO model and the real process leads to deviation between the computed RTO optimum and the true plant optimum. In addition, inconsistency between RTO and MPC models or constraints can result in static offsets and make RTO targets unattainable for the dynamic control layer.
Modifier Adaptation (MA) methods mitigate plant–model mismatch by incorporating cost function gradients into the optimisation problem, allowing convergence toward the true plant optimum despite modelling errors. Ensuring consistency between RTO and MPC, however, is primarily an integration challenge. One integration strategy formulates the MPC economic objective as a function of the distance between RTO solution and the current MPC states [8]. This “soft” integration approach ensures that MPC always prioritises dynamic constraint handling while using remaining degrees of freedom to drive the plant toward the RTO-defined steady-state optimum.
Challenge 2. Frequent variations in economic parameters (product and by-product prices, energy and raw material costs) and transitions between operating modes require RTO to rapidly converge to new optimal conditions. This is often difficult to achieve.
Most RTO technologies rely on steady-state detection for parameter or modifier estimations. The waiting time required to reach steady state for model updates becomes a fundamental limitation, particularly in units with slow dynamics and frequently changing economics or operating conditions. In traditional MA-RTO with many outputs, steady state must be re-established after each Manipulated Variable (MV) excitation to estimate all required gradients. This significantly increases time between successive RTO runs and limits applicability to large-scale optimisation problems.
Directional Modifier Adaptation (DMA) partially alleviates this limitation by estimating gradients only along the most promising directions. A subset of MVs is selected based on objective function sensitivity, reducing the number of required excitations and shortening convergence time.
Challenge 3. High-fidelity first-principles models used in MPA-RTO can introduce numerical robustness and convergence issues. This is particularly relevant for large-scale processes, where optimisation problems may contain hundreds of thousands of equations. Replacing detailed first-principles models with well-trained surrogate models can substantially improve computational efficiency and robustness while preserving essential non-linear behaviour.
Challenge 4. Beyond technical challenges, organisational aspects significantly affect RTO sustainability. Conventional MPA-RTO applications require continuous monitoring and maintenance by skilled engineers with expertise in modelling, control, and optimisation. In many cases, the economic benefits are partially offset by the required engineering resources, making traditional RTO less attractive for small and mid-sized production facilities.
Lightweight, transparent RTO solutions that autonomously adapt to plant changes and are easy to understand and troubleshoot may provide greater practical value for such users than complex proprietary enterprise-grade systems.
1.5 Use of Surrogate Models with Modifier Adaptation
The use of first-principles models for RTO is often impractical due to high computational demand, licensing costs, connectivity limitations, and potential numerical stability issues. Surrogate models offer an alternative approach, where process simulators are used to generate representative datasets that are subsequently approximated using regression or Machine Learning (ML) techniques. In addition to simulated data, historical plant operational data can be incorporated to improve model representativeness. The resulting black-box models can accurately reproduce process behaviour while remaining computationally efficient. Refer to the previous blog post on surrogate modelling [11].
Both the first principles model used to generate the data and the ML model training process can introduce inaccuracies leading to the surrogate model mismatch with the plant. Parametric and structural uncertainty cannot be addressed by changing internal coefficients of the surrogate models because they lack physical interpretability and cannot be related to specific properties of the process.
To compensate for this limitation, the MA technique is applied. Instead of modifying internal model parameters, MA adjusts the optimisation problem itself by introducing zero- and first-order modifiers into the objective function. These modifiers are updated using plant measurements to match the observed gradients and to eliminate offset between predicted and measured economic performance.
The key advantage of MA, as demonstrated in this work, is its theoretical guarantee of convergence to the true plant optimum despite structural plant-model mismatch. When combined with computationally efficient surrogate models, MA provides a practical and robust framework for real-time economic optimisation.
2. The Demo Process: NGL Extraction in an LNG Scrub Column
2.1 Process Overview
The NGL extraction process in the LNG plant Scrub Column unit is used as the case study in this work. A detailed process description, steady-state modelling approach, and the ML methodology used for surrogate model generation were presented in a previous publication [11]. One of the key developments introduced here is the extension of the steady-state model into a dynamic representation of the process. The corresponding process flowsheet of the dynamic model is shown in Fig. 2.1.
The NGL extraction process represents a classical product trade-off problem that is well suited for RTO applications. Operators may choose to increase C5+ recovery and produce larger quantities of condensate or, alternatively, to maximise C5+ content in LNG up to the specification limit, at the expense of reduced condensate production. Each incremental increase in condensate recovery requires progressively higher reflux rates and lower Scrub Column temperatures. Consequently, a larger fraction of NGL is reinjected into the natural gas stream at warmer temperatures, increasing competition for available refrigeration duty and slightly reducing LNG production. This relationship is inherently non-linear; therefore, for every combination of LNG and condensate market prices, there exists a distinct economically optimal level of C5+ extraction.
The Manipulated Variables (MVs) determined by the RTO layer are summarised in Table 2.1. The Disturbance Variables (DVs) serve as inputs to the RTO but are not directly manipulated. These variables are required to improve the accuracy of the economic objective function estimation and are listed in Table 2.2.
Table 2.1 – RTO Manipulated Variables
Manipulated Variable | Description | Operating window |
FY-150 | Scrub Column Recycle ratio | 0.025 – 0.175 |
TC-050 | Scrub Column feed inlet temperature, °C | -20 – 0 |
Process Output Variables (POVs) are intermediate variables required for economic objective function evaluation (Table 2.3). These variables cannot be directly manipulated and instead respond to changes in the MVs. A pre-trained surrogate ML model is used to predict each POV as a function of the MVs and DVs. During optimisation, this surrogate model is evaluated repeatedly to compute the maximised economic objective function value.
Table 2.2 – RTO Disturbance Variables
Disturbance Variable | Description |
FI-050 | Feed flow rate, kg/s |
TI-150 | E-200 outlet (reflux) temperature, °C |
Table 2.3 – Model process variables
Process Output Variable | Description |
FI-250 | Condensate flow rate, kg/s |
QI-250 | Condensate density, kg/m3 |
FI-075 | LNG flow rate, kg/s |
QI-075 | LNG C5+, mol parts |
FI-220 | Long reinjection mass flow, kg/s |
2.2 Economic Objective
The Scrub Column economic objective function is formulated as the net value of products and losses normalised by the feed rate:
where G_CND – value of condensate produced, $/sec;
– G_C5+LNG – value of C5+ comonents in LNG, $/sec;
– G_LNG – LNG losses due to warmer NGL reinjection, $/sec;
- FI-050 – feed flow rate, kg/sec.
Detailed calculations of product values and loss components are provided in [11]. The optimisation problem consists of maximising this objective function in order to increase overall economic profit.
When Modifier Adaptation (MA) is applied, the objective function is augmented as follows [6]:
where FY-150_0 and TC-050_0 denote the reference operating point.
The zero-order modifier ε represents a bias term corresponding to the difference between the plant-measured economic objective and the value calculated using surrogate model predictions of the POVs at the reference operating point.
The first-order modifiers λ_1 and λ_2 represent the differences between the measured objective function gradients and the gradients predicted by the model.
When the plant and model are in perfect agreement, all modifiers are equal to zero. In the presence of plant-model mismatch, non-zero first-order modifiers effectively tilt the objective function surface so that its local gradient aligns with the true plant gradient. This adjustment enables convergence toward the actual plant optimum despite structural or parametric inaccuracies in the surrogate model.
3. ARTData RTO Architechture and Implementation
3.1 Runtime Environment
The RTO demonstration architecture follows a proven pattern previously applied for MPC implementation (Fig. 3.1). The NGL Extraction dynamic process model is implemented in the DWSIM process simulator [12] and runs on a dedicated simulation node connected to a shared network hosting the ARTData server.
The dynamic model communicates with a standalone OPC UA server application running on the same simulation node. Data exchange between the simulation environment and the RTO layer is configured via the OPC UA interface of the ARTData server. Sensor measurements generated by the dynamic process model are continuously transmitted to the ARTData database, while manipulated variable (MV) adjustments are sent in the opposite direction.
The ARTData server data space is organised into three dedicated databases:
dpjt_030_ngl_pcs – the measurement interface layer, including instrument indications, communication health monitoring, and communication fault alarms.
dpjt_030_ngl_est – hosts the RTO algorithm and associated data, including surrogate ML models. The RTO application is implemented and executed within this database. It reads measurements from the measurement interface layer and transmits MV adjustments back to the dynamic process model.
dpjt_030_ngl_aux – auxiliary database used to store multidimensional data generated by the RTO, primarily for visualisation of MV trajectories, POV predictions, and objective function surfaces.
Data from all three databases is presented through dashboard-style visualisation powered by a Grafana server integrated within the ARTData platform. Logs generated by the RTO algorithm are stored in the ARTData file system and can be accessed via a dedicated File Transfer tool.
The dynamic model operates with a virtual time step of 1 second. However, due to computational constraints, the real-time factor is 1/7.5, meaning that 7.5 seconds of real time are required to simulate 1 second of model time. OPC UA data exchange with the model occurs every 5 seconds to ensure that each model time step is communicated.
On the ARTData side, the OPC UA server is scanned every 1 second to collect raw data. The logic within dpjt_030_ngl_pcs, which processes measurements and verifies their health status, executes every 7.5 seconds to align with the effective real-time step of the dynamic model.
The dpjt_030_ngl_est logic executing the RTO algorithm runs every 3 minutes. This corresponds to approximately 24 seconds of virtual dynamic simulation time. Although this represents a relatively high execution frequency for an RTO application, it is justified by the comparatively short transition times of the NGL Extraction model.
The described architecture ensures robustness of logic execution, strong observability through integrated visualisation tools, and high flexibility, allowing additional logic modules or visualisation layers to be incorporated with minimal effort.
3.2 Integration with PCS
The NGL Extraction process base-layer control is implemented within the dynamic process model. The primary control loops include (Fig. 3.2):
Scrub column feed temperature control (E-100 outlet temperature);
Column overhead cooler temperature control (E-200);
Reflux drum level control with output cascaded to reinjection flow rate controller;
Reflux-to-feed ratio control.
In addition, indication instrumentation is configured within the model to collect DVs and POVs measurements.
DVs and POVs measurements are continuously fetched from the dynamic process model and stored and the real-time and time-series database dpjt_030_ngl_pcs. Measurement health status is evaluated based on:
values continuity check (rate-of-change threshold validation);
communication health status based on rolling counter and indicated by XA-100 communication fault alarm.
If measurement raw value exceeds the configured rate-of-change threshold or communication fault is detected then this measurement is assigned bad quality status and the last good quality value is retained.
The RTO application, configured in dpjt_030_ngl_est, retrieves validated historical measurements from dpjt_030_ngl_pcs for steady-state detection, parameter estimation, and optimization.
Integration of MVs with RTO is managed via a watchdog application implemented within the dynamic process model.
The watchdog logic determines whether RTO move requests are passed to the base-layer controllers based on:
MV target and actual modes,
RTO application target and actuals modes.
The RTO application interfaces with MVs exclusively at the controller setpoint level, thereby preserving base-layer control loop integrity. RTO can be decoupled from the base-layer by changing the RTO required mode within the watchdog application.
Refer to Section 2.1 for the complete list of MVs, DVs, and POVs.
3.3 RTO Loop Structure
This section describes the implemented MA RTO sequence and key practical considerations identified during configuration and commissioning. Refer to Fig. 3.3 for the algorithm flow.
3.3.1 Measurement Acquisition
Unlike regulatory control algorithms, RTO requires time-series data rather than instantaneous measurements. Time-series data is required to perform steady-state detection and useful in PV filtering.
The time window length is selected based on the dominant process transition time. In the NGL Extraction dynamic model, process stabilization following a disturbance requires approximately 1.1-2.1 hours of real time (9-17 minutes of simulation time) (Fig. 3.4). A 1-hour window (8 minutes of simulation time) was selected as a compromise between: capturing statistically meaningful variation and avoiding unnecessary delay in steady-state detection.
The retrieved data include:
MV set points readbacks;
PVs corresponding to MVs;
DVs;
POVs;
Historical objective function values.
3.3.2 Steady-State Detection
Steady-state detection is used at multiple stages of the RTO cycle. Each PV is subjected to four independent tests. Each test returns Pass/Fail. Results are encoded as a binary string and converted to decimal representation for diagnostics and troubleshooting (Fig. 3.5). Steady-state is declared only if all applicable tests pass.
3.3.2.1 Data quality
Applied to all time series signals:
Minimal number of samples is 5;
> 80% of data points with Good quality status in the overall window;
> 80% of Good quality in the last 20% of the window;
Final sample quality is Good.
3.3.2.2 Standard deviation
Steady-state requires [8]:
Thresholds are defined based on historically observed steady-state variability.
3.3.2.3 Dickey-Fuller stationarity test
Dickey-Fuller (DF) test evaluates whether the time series exhibits stochastic drift [13]. It relies on fitting a first-order autoregressive model:
and tests:
Stationarity is accepted if the null hypothesis is rejected at the selected significance level (99% confidence level for “constant no trend”). I.e. if the model coefficient for the lagged input is less that 1 then the time-series stabilises at its mean and is considered stationary.
Due to sensitivity of the test to filtering and deterministic signals, DF is applied only to non-deterministic and non-filtered variables, namely POVs, DVs, and PV corresponding to MVs.
3.3.2.4 First-last value match
For MVs, steady-state additionally requires:
This ensures no active movement or operator intervention occurred within the window.
3.3.3 Data Filtering and Reconciliation
Collected raw data contains both random noise and some static offsets. To reduce noise propagation, DVs and POVs are averaged over the last 16 min (2.1 min of simulation time).
A basic mass balance reconciliation is applied:
Feed flow is assumed to be the most accurate measurement. Mass imbalance is redistributed proportionally across product streams.
3.3.4 State Estimation and Bias Update
The surrogate model is evaluated using filtered DVs and current MVs. If steady-state is confirmed for all variables, bias is estimated:
Bias term is then updated using first-order filtering:
When the difference between and becomes less than the standard deviation threshold used for POV steady-state detection, the bias convergence is declared.
After convergence, MA procedure is initiated.
3.3.5 Gradient Mismatch and Modifier Estimation
To compensate structural model mismatch, MA is implemented via automated step-testing.
MA relies on automated plant step-test to evaluate the measured and model-predicted objective function values at different MV values and then find the function gradients of MVs using the step test data. During MA:
bias updates are frozen,
optimisation results are ignored,
MVs are moved accordingly to the predefined step test sequence.
Each MA cycle includes 4 structured moves per MV which minimises drift:
The moves amplitude remains constant. MA requires steady-state detected for MVs, DVs, POVs as well as measured and model-predicted objective function values after each move. After all cycles, steady-state points are approximated using quadratic functions of MVs for:
Measured objective,
Model-predicted objective (with bias applied).
Starting point of the step-test is used as the reference point. For this reference point the fitted quadratic function gradients are estimated analytically with respect to each MV. Resulting gradients mismatch is
To answer if this estimated plant gradient statistically meaningful or just noise, it is needed to evaluate the gradient variance:
Signal to noise ratio is then defined as:
If signal to noise ratio (SNR) is less than one, then gradient mismatch is assumed to be noise and discarded, otherwise it is smoothly scaled until SNR exceeds 3:
Similarly to the bias update, the calculated modifiers are updated using the first order filter:
3.3.6 Optimisation
The surrogate-based objective function is augmented with first-order modifier terms to compensate for plant-model gradient mismatch. For each manipulated variable u_i the distance from the modifier reference point is:
At u_i = u_i,0 the modifier contribution is zero and the objective function coincides with the model-based objective.
To reflect decreasing confidence in locally estimated gradients as the operating point moves away from the reference, a decay factor is introduced:
where δ_i is the local MV move limit;
κ is the local MV move limit.
The modified objective function becomes:
where λ_(i,used) are filtered gradient modifiers. This formulation implements first-order modifier adaptation with distance-based trust-region decay.
The optimisation is performed in two stages:
1. Global search. Bayesian optimisation is applied over the full MV operating window to identify a global optimum.
2. Local steps. Optimisation is performed using MV step limits δ_i locally over a series of steps.
The local step-by-step optimisation limits potential aggressive moves and allows direct integration of RTO with the base-layer control without use of MPC. For convex objective functions, the local refinement converges to the global optimum. For non-convex problems, the two-stage approach may result in convergence to a poor local optimum, but can be detected by comparing the prediction series with the global optimum results.
Bayessian optimisation is used as an optimisation algorithm. To avoid insignificant or oscillatory adjustments, MV moves are suppressed when the MVs are within move size limit from the global optimum:
Bayesian optimisation acquisition function parameters are tuned to favour exploitation during local steps stage to reduce unnecessary exploration.
3.3.6 Application of MV Moves
The local optimisation generates a prediction sequence of MVs, POVs and objective values across the move horizon. Analogous to a receding-horizon approach, only the first move in the predicted sequence is implemented. If the optimisation sequence is unavailable (e.g., internal error or RTO inactive), the last MV readback is retained.
On the dynamic process model side, the watchdog application receives MV statuses, proposed MV moves, RTO operational mode. If both RTO and MV are active, the watchdog forwards the move to the corresponding controller setpoint. Otherwise, the move is blocked.
3.4 Operational Modes
To meet various operational requirements RTO was developed with following modes:
INITIALISE
Purpose: quick model alignment during start-up.
Bias is updated instantaneously:
No steady-state detection is performed.
No modifier adaptation.
No optimisation.
No MV moves applied.
OBSERVE
Purpose: monitoring and validated bias tracking.
Surrogate models evaluated continuously.
Steady-state detection enabled.
Bias updated only when steady-state is confirmed.
Bias updated using first-order filtering.
No modifier adaptation.
No optimisation.
STANDBY
Purpose: semi-automatic optimisation with human in the loop.
Same behaviour as OBSERVE.
Global and local optimisation executed.
Prediction sequence of MVs, POVs, and objective values generated.
No MV moves sent to plant.
ACTIVE
Purpose: full MA-RTO operation.
Same as ACTIVE.
Periodic MA step-test performed.
Bias updates frozen during MA step-test.
First-order modifiers updated after successful MA step-test completion.
Optimisation uses objective function augmented with the updated modifiers.
Additional control features include:
individual MV On/Off mode switching. Disabled MVs are excluded from optimisation and their values remain fixed.
reset of the modifiers to 0 via MA RESET switch. Optimisation reverts to pure model-based objective after the reset and new MA cycle is needed to restore the gradients correction.
logging of RTO algorithm critical exceptions and MA routine details to facilitate troubleshooting.
3.5 HMI Visualisation and Monitoring
The MA-RTO HMI within ARTData is implemented using two dedicated Grafana dashboards to maintain clear separation between base-layer control monitoring and supervisory optimisation functions:
ARTData NGL PCS – enables NGL Extraction process and base-layer performance monitoring. It shows communication status, measurements quality status, base-layer controller key metrics (MV, PV, SV) and plant measurements.
ARTData NGL RTO – provides access to RTO modes and switches, shows prediction series for MVs and POVs, enables monitoring of bias and modifiers updates, shows steady-state detection statuses, objective function trends and 3D surface of the objective with and without MA.
The HMI walkthrough is provided in the following slides:
4. ARTData RTO Operation and Results
4.1 Base Case Scenario
In the base case scenario, the MA-RTO is deployed for the NGL Extraction dynamic model where the process behaviour and performance are identical to the steady-state process model used in the RTO surrogate models training. While there are minor inaccuracies in the ML models predictions, the experimental plant optimum P is located close to the model optimum M, estimated from the surrogate models (Fig. 4.1). I.e. in the base case scenario there is no significant structural uncertainty of the RTO underlying surrogate model.
During the performance test, the RTO was activated at the starting operating point S positioned far from the optimum. It is left in INITIALISE mode until process is in steady-state and then switched to ADAPTIVE mode. Once active, the RTO quickly moves the operating point towards the surrogate model optimum M through a series of incremental steps applied to both TC-050 and FY-150 MVs. The trajectory of moves is driven by the highest gradient direction at each step. Because deadband is applied on the moves near the RTO estimated global optimum to avoid chasing minor fluctuations, the RTO stops at the model-driven best point O closely located to M.
At the point O steady-state condition is reached and RTO application updates the model bias until model predictions match plant POVs within specified standard deviation thresholds. Bias update does not result in the surrogate model optimum M shift or model-driven best point O change.
After the objective function stabilises at point O, MA algorithm activates. Over 5 successive MA iterations (Table 4.1) the estimated modifiers shift operating point to MA driven best point Q located closer to the experimental plant optimum P.
Table 4.1 – MA iterations in base case scenario
MA Iter | MV | Measured Objective Gradients | Model Objective Gradients | Gradient Difference | Gradient Standard Deviation | Calculated modifiers |
1 | FY-150 | 7.1129 | 5.7130 | 1.3999 | 1.2133 | +0.2070 |
TC-050 | -0.0299 | -0.0154 | -0.0145 | 0.0121 | -0.0023 | |
2 | FY-150 | 5.3827 | 5.7642 | -0.3815 | 1.3556 | -0.0000 |
TC-050 | -0.0485 | -0.0156 | -0.0329 | 0.0135 | -0.0216 | |
3 | FY-150 | 1.5143 | 3.9231 | -2.4088 | 1.3703 | -0.8270 |
TC-050 | -0.0721 | -0.0110 | -0.0611 | 0.0137 | -0.0611 | |
4 | FY-150 | 5.3201 | 0.0932 | 5.2268 | 1.3322 | 5.2268 |
TC-050 | -0.0072 | 0.0090 | -0.0162 | 0.0133 | -0.0026 | |
5 | FY-150 | 1.7471 | 1.2238 | 0.5232 | 1.5515 | -0.0000 |
TC-050 | -0.0440 | 0.0032 | -0.0473 | 0.0155 | -0.0472 |
With the calculated modifiers being impacted by the signal noise, it is important to apply filtering when updating the used modifier values (Fig. 4.2).
4.2 Adaptation to Altered Plant Performance
The base case scenario assumes constant absorber tray efficiency at 35%. In real columns the efficiency drops when amount of reflux creates liquid flow exceeding flooding limit. The altered plant case scenario adds this phenomenon into the dynamic process model.
While in the base case scenario reflux increase always leads to objective function increase, hence optimal point is as the FY-150 MV boundary 0.1750, in the altered plant the objective function peaks at FY-150 ~0.1000 and then reduces with reflux ratio increase.
5. Key Takeaways
5.1 Why MPC Alone Is Not Enough for Economic Optimisation
MPC is based on simplified linear models describing relationships between MVs, DVs and controlled variables, and it is primarily aimed at solving dynamic control problems. It is particularly effective in constrained control applications because constraints are natively handled by the MPC optimisation algorithm [4].
If the plant optimum lies on active constraints and the linear process gains remain sufficiently accurate in that region, MPC can drive the process towards that optimal point. However, when the true plant optimum is unconstrained, the process gains between MVs and the objective function inherently change sign when passing through zero at the extremum.
MPC is not designed to search for a steady-state economic optimum, nor to adapt nonlinear model gradients. Both capabilities are essential elements of RTO.
5.2 Can RTO Work Without MPC
In cases where a process unit is relatively small and does not exhibit strong process interactions, base-layer PID control may be sufficient to handle all dynamic control requirements, and MPC may not provide additional economic benefit.
If an economic optimisation problem still exists, RTO can be implemented directly on top of the base-layer control, provided that:
RTO moves are properly managed, and
Base-layer constraints are reflected in the RTO optimisation formulation.
This NGL Extraction RTO demo case demonstrates that such integration is feasible.
For larger units and plants where MPC is economically justified, MPC naturally becomes the integration layer between RTO and the base-layer control system.
5.3 When RTO Based on Surrogate Models Makes Sense
Surrogate models serve as replacements for first-principles models when the latter are:
not readily available,
not accurate enough,
missing programmable interfaces,
unreliable in a PCS environment,
too costly to develop and maintain.
Surrogate models can be built using:
data generated from first-principles models,
historical plant data, or
a hybrid combination of both.
When surrogate models are used in RTO, MA layer becomes necessary. MA performs best when:
process economics are clear (i.e., good SNR),
the number of MVs is limited,
the process transition time is relatively short.
5.4 Why Modifier Adaptation Is Not Optional
RTO without MA retains only bias update functionality, which reduces parametric uncertainty. However, if the true plant optimum differs from the model-based optimum due to structural uncertainty, such limited RTO implementation cannot converge to the true optimum.
With classical MPA-RTO, the underlying first-principles model can be continuously maintained and improved to reduce structural mismatch. With ML-based surrogate models, such incremental structural improvement is difficult due to their black-box nature.
MA addresses this limitation in both cases by testing live plant responses and augmenting the objective function with modifier gradients for each MV. This ensures that the modified objective function optimum converges towards the plant optimum rather than the model optimum.
5.5 What to Focus on First in Real Projects
In real projects, technical and organisational feasibility should be assessed as early as possible. For MA-RTO based on surrogate models, key questions include:
Is the process economics clear and sufficiently sensitive to process conditions? Why cannot the optimisation function be embedded within the existing MPC layer (e.g., due to unconstrained or nonlinear behaviour)?
Is there a reliable and accurate steady-state first-principles model? Can it be used to generate training data for ML models?
Is there sufficient historical operating data with adequate variability to validate ML model accuracy?
What are the required MV move sizes and stabilisation periods to reliably estimate objective function gradients? Are these moves acceptable from an operations perspective? What is the probability of disturbances occurring during the stabilisation window and corrupting gradient estimation?
Are the base-layer and MPC layers ready for integration with RTO? What are the RTO decision variables and where should they be integrated?
What are the intended deployment stages (e.g., model-based optimisation only, open-loop advisory mode, full MA-RTO)?
What is the long-term maintenance and monitoring plan?
5.6 What Would Change in Full-Scale Deployment
MA provides significant benefits to RTO, but it is also the most complex part of the overall algorithm. It relies heavily on steady-state detection (SSD), and a broader set of SSD methods would likely be required in a real industrial deployment.
Furthermore, historically estimated MA gradients linked to their respective trust regions could be reused more effectively when updating modifiers via first-order filtering. In the current implementation, previously used modifier values are retained regardless of their trust location.
Abbreviations
Abbreviation | Full description |
ARTData | Advanced Real-Time Data Processing Platform |
DCS | Distributed Control System |
DF | Dickey-Fuller Test |
DMA | Directional Modifier Adaptation |
DV | Disturbance Variable |
ERP | Enterprise Resource Planning |
HMI | Human Machine Interface |
LNG | Liquefied Natural Gas |
MA | Modifier Adaptation |
MES | Manufacturing Execution System |
ML | Machine Learning |
MPA | Model Parameter Adaptation |
MPC | Model Predictive Control |
MV | Manipulated Variable |
NGL | Natural Gas Liquid |
OPC UA | Open Platform Communications Unified Architecture |
PCS | Process Control System |
PID | Proportional Integral Derivative controller |
PLC | Programmable Logic Controller |
POV | Process Output Variable |
PV | Process Value |
RTO | Real-Time Optimisation |
SNR | Signal to Noise Ratio |
SSD | Steady-State Detection |
SV | Set-point Value |
VSD | Variable Speed Drive |
References
[1] S. Skogestad. Chapter 19 Real-Time Optimization.
[2] Alejandro G. Marchetti. Modifier Adaptation for Real-Time Optimization – Methods and Applications. Processes.
[3] ISA-95 Series of Standards.
[4] Profware Systems. End-to-End MPC Implementation with ARTData: From Process Model to Operator Interface Using Gekko and TCLab.
[5] B. Chachuat, B. Srinivasan, D. Bonvin. Adaptation strategies for real-time optimization.
[6] A. G. Marchetti, G. François, T. Faulwasser, D. Bonvin. Modifier Adaptation for Real-Time Optimization - Methods and Applications.
[7] Jose Eduardo Alves Graciano. Real-Time Optimisation in Chemical Processes: evaluation of strategies, improvements and industrial application.
[8] P. A. Delou. A Real-Time Optimization Strategy for Small-Scale Facilities
and Implementation in a Gas Processing Unit.
[9] Dinesh Krishnamoorthy, Sigurd Skogestad. Real-Time optimization as a feedback control problem – a review.
[10] Yuanyuan Zou, Xu Ma, Yaru Yang, Shaoyuan Li. An overview of chemical process operation-optimization under complex operating conditions.
[11] Profware Systems. Machine Learning for Process Control. Part 2: NGL Extraction Unit Optimisation with Surrogate Models.
[12] DWSIM chemical process simulator.
[13] Evren Mert Turana, Johannes Jaschkea. A simple two-parameter steady-state detection algorithm: Concept and experimental validation.
[14] Ken McCarley. Finding the Capacity of a Distillation Column In Section: Reactions and Separations.
Comments