Optimization Algorithms Advance
Contents
01 Why APC Needs Optimization
04 Beyond Linear MPC — Advanced Solvers
02 Core Algorithm: Quadratic Programming
05 Five Most Popular APC Applications
03 Real-Time Optimization (RTO)
06 Choosing the Right Algorithm
01 —
Why APC Needs Optimization
Welcome to the Optimization Algorithms Advanced Process Control. Every industrial process is a balancing act between competing objectives and hard physical limits. A refinery wants to maximize the yield of high-value diesel while minimizing fuel consumption — but it must simultaneously keep column pressures within safe bounds, avoid flooding trays, and maintain product specifications. A single-loop PID controller, no matter how well tuned, cannot balance these competing demands at once. It controls one thing at a time and has no awareness of constraints elsewhere in the system.
Advanced Process Control solves this by framing the control problem as a mathematical optimization problem — find the best set of control actions that minimize (or maximize) an objective while satisfying all constraints. This is not a new idea; operations research has studied such problems since the 1950s. What changed in the 1980s and 1990s was the availability of fast computers capable of solving these problems in the seconds or minutes available between control execution cycles in a live industrial plant.
The Central Idea
“APC optimization is not about finding a perfect answer — it is about finding the best achievable answer, given real physical limits, in the time available between one control cycle and the next. Speed and feasibility matter as much as mathematical elegance.”
The optimization problem in APC has three key ingredients: an objective function (what to maximize or minimize — typically economic value, energy use, or deviation from targets), decision variables (the control moves the system can make — valve positions, flow rates, temperatures), and constraints (hard limits on equipment safety, soft limits on product quality, and rate-of-change limits on actuators). The choice of optimization algorithm depends directly on the mathematical character of these three components.
02 —
Core Algorithm: Quadratic Programming in MPC
The dominant optimization algorithm inside commercial Model Predictive Control (MPC) platforms is Quadratic Programming (QP). QP solvers are the workhorses of industrial APC — used inside AspenTech DMC3, Honeywell Profit Controller, Shell SMOC, and virtually every other commercial MPC package. Understanding why QP is so well-suited to the MPC problem requires looking at the mathematical structure MPC creates.
When an MPC controller uses a linear dynamic model to predict process behavior, the relationship between future control moves (the decision variables) and future controlled variable values (what we want to manage) is linear. If the objective function is then expressed as the sum of squared deviations from targets — a quadratic function — the resulting optimization problem is a QP: a quadratic objective subject to linear constraints. QP problems have a unique global minimum and can be solved extremely efficiently using algorithms like the Active Set Method or Interior Point Methods.
How the QP is Structured Inside MPC
Objective Function — Minimize Deviation and Move Size
The QP objective minimizes a weighted sum of two terms: the squared deviation of predicted controlled variables (CVs) from their set-points or constraint limits over the prediction horizon, and the squared magnitude of control moves (MV changes). The move-suppression weight is a critical tuning parameter — large values produce smooth, conservative control; small values allow aggressive, fast responses.
Inequality Constraints — Enforcing Physical Limits
Linear inequality constraints enforce high and low limits on MVs (valve position limits, flow rate limits), CV operating ranges (column temperature ranges, product spec limits), and MV rate-of-change limits (how fast a valve can move per control cycle). Hard constraints must never be violated. Soft constraints are handled with slack variables — the optimizer can violate them at a penalty cost when necessary to maintain feasibility.
Steady-State Target Layer — Economic Optimization
Many commercial MPC packages split the QP into two sequential problems: a steady-state target calculation (SSTC) that finds the economically optimal operating point within the feasible region, and a dynamic move calculation QP that drives the process toward those targets as smoothly as possible. The SSTC is typically a Linear Program (LP), which is even simpler and faster to solve than a QP.
A modern QP solver can handle MPC problems with 50 manipulated variables, 100 controlled variables, and a 60-step prediction horizon — thousands of decision variables and constraints — in under 100 milliseconds on standard industrial hardware. This speed makes real-time implementation practical, even for fast processes.
03 —
Real-Time Optimization (RTO) — Nonlinear Programming
MPC with QP is powerful, but it has an important limitation: it uses a linear process model. Linear models are excellent approximations near a given operating point, but many industrial processes are highly nonlinear — the relationship between, say, reactor temperature and reaction rate is exponential (Arrhenius equation), and a linear model captures this accurately only in a narrow range. For plant-wide economic optimization that must operate across wide ranges, a more powerful approach is needed.
Real-Time Optimization (RTO) operates one layer above MPC and uses a rigorous nonlinear steady-state model of the entire process — built from first principles or from identified nonlinear correlations — to compute economically optimal set-point targets. Because it uses nonlinear models, the resulting optimization problem is a Nonlinear Program (NLP), solved by algorithms such as Sequential Quadratic Programming (SQP) or Interior Point NLP solvers like IPOPT.
SQP works by solving a sequence of QP subproblems, each of which approximates the NLP locally using a quadratic model of the objective and a linear model of the constraints. This iterative approach converges to the NLP optimum and is the algorithm of choice in tools like AspenTech’s RTO+ and Honeywell’s RMX. RTO execution cycles are typically 30 minutes to 4 hours — far slower than MPC — because the NLP solve is more computationally intensive and requires the plant to reach approximate steady-state before the solution is meaningful.
RTO vs MPC — Algorithm Comparison
MPC (QP / LP)
— Linear dynamic process model
— Execution: every 1–5 minutes
— Optimizes control moves over horizon
— Handles dynamics & constraints
RTO (SQP / NLP)
— Rigorous nonlinear steady-state model
— Execution: every 30 min – 4 hours
— Optimizes economic targets for MPC
— Handles plant-wide economics
04 —
Beyond Linear MPC — Advanced Optimization Solvers
While QP-based linear MPC dominates industrial practice, several specialized optimization algorithms address problems where linearity assumptions break down or where the nature of the process demands different mathematical tools.
Nonlinear MPC (NMPC)
Uses a nonlinear dynamic model directly inside the MPC prediction and optimization loop. The resulting NLP is solved at every execution cycle using SQP or Multiple Shooting methods. NMPC is used for highly nonlinear batch reactors, polymerization processes, and biological systems where a linear model cannot represent the process adequately even over a short horizon.
Genetic Algorithms (GA)
Population-based evolutionary optimization that evaluates many candidate solutions simultaneously, selecting, crossing, and mutating them toward better objective values. GAs do not require gradients and handle discontinuous, non-convex, and combinatorial problems well. Used in APC for process scheduling optimization, catalyst change scheduling, and furnace pass-balancing problems where the search space has many local optima.
Particle Swarm Optimization (PSO)
A swarm-intelligence algorithm where candidate solutions (particles) move through the search space, influenced by their own best-found position and the global best. PSO is computationally light and handles multimodal, noisy objective functions. It appears in APC applications involving neural-network-based process models, where the model’s nonlinearity makes gradient-based QP impractical.
Stochastic / Robust MPC
Extends classical MPC by explicitly modeling uncertainty in the process model and disturbances. The optimization problem becomes a stochastic program — minimize expected cost while ensuring constraint satisfaction with a specified probability. Used in energy systems, grid-connected storage, and chemical plants with highly variable feed compositions. Solved via scenario trees or chance-constrained programming formulations.
The practical selection among these algorithms depends on process nonlinearity, execution time requirements, availability of gradient information, and the commercial platform in use. For the vast majority of continuous industrial processes, linear MPC with QP remains the gold standard due to its speed, reliability, and the maturity of supporting software tools.
Five Most Popular APC Optimization Applications
APP 01
Crude Distillation Unit (CDU) Optimization
Algorithm: LP Steady-State Optimizer + QP Dynamic MPC
The crude distillation unit is the entry point for all refinery operations and the single largest lever on refinery profitability. Its APC application is among the most mature and widely deployed in the world, with implementations at virtually every major refinery globally. The optimization objective is to maximize the economic value of the product slate — shifting yield toward higher-value cuts (diesel, kerosene) and away from lower-value fuel oil — while satisfying all product specification constraints and equipment operating limits.
The LP steady-state optimizer solves for the economically optimal product draws and operating conditions using current product prices (fed from the plant’s planning system). These targets are passed to a QP-based MPC that dynamically steers 20–40 manipulated variables — including preheat train flows, stripping steam rates, reflux ratios, and pump-around duties — toward those targets while managing dozens of temperature, pressure, and quality constraints in real time.
Typical Benefit: $3–15 million/year per CDU. Yield shift to high-value products of 0.5–2.0 vol%, energy reduction of 3–8%, CV variability reduction of 40–60%.
APP 02
Fluid Catalytic Cracker (FCC) Optimization
Algorithm: SQP Nonlinear RTO + QP MPC
The Fluid Catalytic Cracker converts heavy gas oil into gasoline, LPG, and other light products using a circulating catalyst system operating at extreme temperatures (500–550°C in the reactor, 680–730°C in the regenerator). The FCC is one of the most complex and nonlinear processes in a refinery — catalyst circulation rate, regenerator air rate, reactor temperature, and feed preheat all interact in ways that a linear model cannot fully capture. This makes the FCC one of the canonical applications for nonlinear RTO.
An SQP-based RTO uses a first-principles kinetic model of the FCC to compute optimal reactor severity and catalyst-to-oil ratio that maximize gasoline and LPG yield for the current feed slate and product prices. These targets update every 1–2 hours. An underlying QP MPC manages the real-time dynamics — reacting to feed rate changes, catalyst activity fluctuations, and regenerator air blower constraints within minutes — keeping the unit on the optimal trajectory set by the RTO.
Typical Benefit: $5–20 million/year per FCC. Gasoline yield increase of 1–3 vol%, coke yield reduction, regenerator temperature stability, and extended catalyst life.
APP 03
Ethylene Plant Cracking Furnace Optimization
Algorithm: NMPC + Genetic Algorithm (Decoking Scheduling)
Ethylene production begins in cracking furnaces that thermally crack hydrocarbon feeds (ethane, naphtha, LPG) at temperatures of 780–870°C. A world-scale ethylene plant typically operates 6–12 cracking furnaces simultaneously, and optimizing across the entire furnace bank is one of the most computationally rich APC applications in the petrochemical industry. The optimization has two distinct timescales: real-time furnace control (NMPC, executing every minute) and decoking schedule optimization (determining which furnace to take offline for coke removal and when, executed over a 24–72 hour horizon).
NMPC manages coil outlet temperature, steam-to-hydrocarbon ratio, and feed distribution across pass inlets — all highly nonlinear due to the cracking kinetics. The decoking scheduler uses a Genetic Algorithm because the problem is combinatorial (which furnace to decoke, and in what order, to maximize overall ethylene output while meeting demand) — a class of problem where GA outperforms gradient-based methods by exploring the discrete solution space without getting trapped in local optima.
Typical Benefit: $8–25 million/year per ethylene plant. Ethylene yield improvement of 0.3–1.2 wt%, furnace run-length extension of 10–20%, specific energy reduction of 2–5%.
APP 04
Power Plant Load Dispatch and Heat Rate Optimization
Algorithm: LP Economic Dispatch + QP MPC + Stochastic Optimizer
In power generation, optimization algorithms serve two distinct functions. At the plant level, heat rate optimization — minimizing fuel consumption per megawatt-hour — is an MPC problem. The QP-based APC controller manages burner configuration, excess air, steam temperature and pressure, feedwater flow, and combustion air distribution to maximize thermodynamic efficiency while following the load set-point dispatched by the grid operator. The objective function is a linearized heat rate curve; the constraints include environmental emission limits (NOₓ, SOₓ, opacity) and equipment protection limits.
At the multi-unit level, economic load dispatch uses a Linear Program to allocate total generation demand across multiple generating units (each with its own nonlinear heat rate curve, approximated as piecewise linear). The LP minimizes total fuel cost across the fleet for a given total output requirement — a classic resource allocation problem. Modern implementations use Stochastic MPC when renewable generation is integrated, explicitly modeling wind and solar uncertainty in the optimization to ensure grid stability with minimum reserve cost.
Typical Benefit: Heat rate improvement of 1–3% (worth $2–8 million/year on a 500 MW plant). NOₓ emissions reduction of 15–30%. Load following speed improvement of 20–40%.
APP 05
Mineral Processing — Grinding Circuit Optimization
Algorithm: PSO / Fuzzy-MPC + QP MPC
In mining and mineral processing, grinding circuits — the combination of SAG (Semi-Autogenous Grinding) mills and ball mills — are responsible for reducing ore to the fine particle sizes needed for downstream flotation and recovery. Grinding is typically the single largest energy consumer in a mine, often representing 30–50% of total site electricity consumption, and it directly determines metal recovery in subsequent processing steps. Optimizing the grinding circuit is therefore among the highest-value APC applications in the mining industry.
The optimization challenge is significant: ore hardness varies continuously and unpredictably with the mining face, and the target product particle size (P80) is typically measured only by periodic sieving — not in real time. A PSO-based or Fuzzy-MPC optimizer handles the outer loop, estimating ore hardness from power draw signatures and optimizing mill speed, feed rate, and water addition to maximize throughput at target grind size. An underlying QP MPC handles the fast dynamics — mill load, cyclone pressure, sump level — that must be controlled second-by-second to protect equipment.
Typical Benefit: Throughput increase of 3–8%, energy consumption reduction of 4–10%, improved grind consistency leading to 1–3% improvement in downstream metal recovery.
06 —
Choosing the Right Algorithm
The optimization algorithm used in any given APC application is not chosen arbitrarily — it follows directly from the mathematical properties of the process and the practical requirements of the implementation. For continuous processes near a single operating point, linear MPC with QP provides fast, reliable, and globally optimal solutions. When the process is highly nonlinear over the operating range, NMPC with SQP or a nonlinear RTO layer provides the accuracy needed to capture the true economic optimum.
For scheduling, combinatorial, and multi-modal problems where gradient information is unavailable or misleading, evolutionary methods like Genetic Algorithms and Particle Swarm Optimization provide robustness at the cost of computational speed. And when uncertainty in feedstocks, demand, or renewable generation is a first-class concern, stochastic programming formulations explicitly trade off average performance against worst-case risk — a fundamentally different objective from the deterministic optimizers that dominate classical APC.
Quick Algorithm Selection Guide
Use QP / LP when…
Process is approximately linear near operating point, fast execution (<1 min) is required, commercial MPC platform is available, and constraints are well-defined linear inequalities.
Use SQP / NMPC when…
Process has strong nonlinearities (reactors, bioreactors, furnaces), operating range is wide, gradient information is available, and accuracy of the optimal solution is critical to economics.
Use GA / PSO when…
Problem is combinatorial or scheduling-based, objective has many local optima, gradient is unavailable (black-box neural network models), and execution time of minutes to hours is acceptable.
Use Stochastic MPC when…
Process faces significant uncertainty in feed, demand, or environment (renewable energy, variable ore grade), constraint violations carry severe consequences, and risk management is as important as average performance.
<100ms
QP solve time for typical industrial MPC problem
5–20×
Typical ROI multiple achieved within first year
90%+
Market share held by QP-based MPC in continuous process APC
The future of APC optimization points toward hybrid approaches — combining the speed and reliability of QP-based MPC for real-time control with the economic depth of nonlinear RTO, the robustness of evolutionary algorithms for scheduling, and the uncertainty-awareness of stochastic programming for operations in a more volatile world. The algorithms are maturing rapidly, but the fundamental insight remains unchanged: every dollar earned from industrial optimization traces back to the same act — solving a well-posed mathematical problem, faster than the process can move.
