Closed-loop AI process control & real-time optimisation for continuous plants

Continuous plants are run by PID loops and model-predictive control (MPC/APC) that hold setpoints, while operators and engineers decide the setpoints themselves. Closed-loop AI goes one step further: a learned process model or a reinforcement-learning (RL) agent continuously moves setpoints on the DCS to push yield, throughput and energy efficiency toward the economic optimum, inside hard operating limits. Safety instrumented systems remain independent and untouched. Benefits are plant-specific and are mostly reported by vendors; margin and energy gains of a few percent matter a lot at refinery or kiln scale.

Problem & value

• Today: units run on PID plus linear MPC (e.g. DMCplus, Profit Controller). Linear MPC models degrade as catalysts age, feeds change and equipment fouls; re-identification is expensive. Many difficult units (distillation with strong disturbances, kilns, mills) are still partly hand-steered by experienced operators, so performance varies by shift.
• AI value: nonlinear, continuously retrained models adapt to changing feed and fuel quality; the agent acts every few seconds or minutes, consistently across shifts; savings in fuel, steam, power and emissions (relevant for EU ETS cost exposure) and fewer off-spec batches.
• Evidence: Yokogawa and NAIST's FKDPP RL algorithm ran a distillation column at an ENEOS Materials plant in a 35-day trial in 2022 and then for almost a year, and was formally adopted for direct control in 2023; the same technology was applied at Aramco's Fadhili gas plant with vendor-reported preliminary reductions of 10-15% in amine and steam use and about 5% in power (2025) [1][2][3]. These are vendor claims; independent, peer-reviewed benefit data are scarce.

Variants

• Refining and petrochemicals: crude/vacuum units, FCC, hydrocrackers, ethylene crackers; hybrid models combining first principles with deep learning (Imubit "Closed Loop AI Optimization").
• Chemicals and specialities: reactors, distillation columns, utilities; batch-to-continuous.
• Cement, lime, glass: kiln, raw mill, cooler and alternative-fuel optimisation (stable burning zone, lower specific heat consumption, higher alternative-fuel share).
• Advisory-first ("open loop") variant: same models, but recommendations are shown to operators who accept them; the standard first stage before write-back.
• Utilities and steam networks: boiler/CHP load allocation, see OPT-023.

How it works

1. historian + analysers + LIMS
2. data cleaning, tag mapping, regime detection
3. process model (hybrid first-principles + deep learning, or simulator-trained RL environment)
4. optimiser / RL policy (objective: margin, energy, emissions; constraints: quality, limits)
5. safety envelope & rate limiter (hard clamps, validity checks, model-health watchdog)
6. OPC UA / vendor connector write to DCS setpoints (SP of existing regulatory loops)
7. plant response monitored
8. retraining / policy update ↘ operator HMI: enable/disable, bounds, explanations, one-click fallback to PID/APC

• AI components: (1) models of dynamics (neural nets, hybrid models, state-space models); (2) RL agents trained offline on a simulator or on historical data (offline/conservative RL) and deployed with limited action range; (3) classic MPC or solvers for constraint handling, often wrapped around the learned model.
• Safety architecture: the AI writes only to setpoints of existing basic control loops, never to final elements directly; alarms, interlocks and emergency shutdown remain in an independent safety instrumented system per IEC 61511 / IEC 61508, which the AI cannot bypass [5]. Cybersecurity zones follow IEC 62443; write access is one-way through a controlled gateway.
• Human-in-the-loop: operators can switch the AI off at any time; engineers set bounds and approve model updates under management-of-change; deviations trigger automatic fallback.

Data preparation

The data lives in the plant historian (PI, IP.21, PHD), the DCS and PLCs, the LIMS for lab results, online analysers, maintenance records in CMMS or SAP PM, and sometimes in operator logbooks on paper or in spreadsheets. Process engineers often hold the decisive knowledge about operating modes in their heads.

Typical gaps: tags without consistent names or units, sensor drift and frozen values, lab results with irregular timing and delay, undocumented grade changes, shutdowns and manual interventions that distort training data, and an absence of labelled disturbances. Access to the OT network is restricted and historian licences may limit exports.

Preparation steps: (1) inventory tags, sources and owners (PREP-001, PREP-005); (2) clean and align time series: resample, mark outages, correct drift, align lab values to process time; (3) build a tag dictionary with units and meaning (PREP-012, PREP-015); (4) annotate operating modes and events with process engineers (PREP-020, PREP-016); (5) build a test set and a dynamic simulator or surrogate model for validation (PREP-018); (6) set up a read-only data mirror in an IT/OT demilitarised zone. Use Python, Polars, TimescaleDB or the historian's own tools. Plan for 50 to 60 percent of project effort here, so the rating is high.

Implementation paths

Which path for whom

• Enterprise (plant operator): C, because process and recipe data are trade secrets, the write path to the DCS belongs inside the OT network, and one on-premises platform (historian, simulator, solver) can serve several units; vendor suites are a valid alternative for operators who want a supported product.
• Developer / integrator: C, because open-source RL, simulation and solver tools allow building and validating agents against a unit simulator before any DCS connection.

Path A — Buy (off-the-shelf)

• Vendors: Yokogawa (Autonomous Control AI based on FKDPP, Japan), Imubit (US; closed-loop AI optimisation for refining/chemicals), AspenTech (Aspen DMC and AI-driven hybrid models, now under Emerson), Honeywell (Forge / Profit Suite), ABB (Ability Expert Optimizer, strong in cement and minerals; Swiss/Nordic engineering), Siemens (EU; process control and AI add-ons for PCS neo), Emerson, AVEVA, Rockwell and Schneider Electric (EU). Cement-specific: FLSmidth (EU, Denmark) and Holcim/Heidelberg Materials' own digital programmes are public references. Vendor claims must be validated on your plant through a staged pilot.
• Choose A when: you have a historian with years of clean data, a DCS vendor ecosystem and no in-house RL team. Plan for site-level engineering effort and a multi-month advisory phase.

Path B — Build on cloud APIs

• Cloud is used for training and analytics, not for the control loop: Azure, AWS or Google historian ingestion (AVEVA PI, Seeq, AWS IoT SiteWise), training on EU-region GPU nodes, model packaging to the edge. LLMs help operators with explanations and shift-report summaries but never issue setpoints.
• Choose B for the analytics layer when central data science teams build models for several sites and the actuation runs on-premises.

Path C — Self-host (open source)

• Stack: OPC UA clients (open62541, asyncua), Python with PyTorch, Stable-Baselines3 / RLlib or d3rlpy (offline RL), do-mpc and CasADi / Pyomo / IPOPT for model-based control, Modelica / OpenModelica or DWSIM as simulators, MLflow, Kafka. Control-in-the-loop runs on industrial edge PCs (HW1); training fits HW2-HW3.
• Ops effort: needs process-control, OT-security and ML engineers; certifiable fallback logic is entirely your responsibility.

Path D — Custom training

Where the value is: train the plant-specific dynamics model or RL policy against a calibrated simulator or digital twin. Needs 1-3 years of tag history with step tests or natural excitation, validated first-principles models, and a sim-to-real validation plan (shadow mode, then bounded write). Offline RL from logs is attractive because live exploration on a real plant is unacceptable.

State of the art (2026)

• RL moves from papers to formal plant adoption in a few documented cases (FKDPP at ENEOS Materials; later gas-processing deployments), usually on well-instrumented units with strong simulators or abundant data [1][2][3].
• Hybrid models (physics-informed neural nets, differentiable simulators) dominate commercial offers because they extrapolate better and are easier to validate than pure black boxes.
• Safe RL techniques: action bounding, constrained policy optimisation, shielding and offline RL with conservative value estimates; runtime monitors check that inputs stay in the training distribution.
• Agentic and LLM layers are emerging for operator copilots, but remain outside the control path.
• Limits and failure modes: distribution shift (new feed, sensor drift, fouling), bad or frozen sensors, reward mis-specification (agent exploits a lab-measurement lag), poor sim-to-real transfer, operator distrust leading to switch-offs, and benefit claims measured against flawed baselines. Always measure with A/B periods or matched-baseline tests.

Shared resources & combinations

• Same time-series/historian pipeline as PRED-014 (soft sensors feed the controller), OPT-009 (offline parameter tuning), OPT-024 (water treatment), ANOM-005 (process anomaly detection).
• Simulation and sim-to-real: shared with ROBO-010; solver layer shared with OPT-018 and OPT-023.
• Energy side: OPT-007 (building/HVAC) and OPT-023 (CHP) for utilities scheduling around the same plant.
• Hardware: edge industrial PC at the plant (HW1), shared GPU server for training (HW2/HW3).

Effort, cost & KPIs

• Path A: typically 6-18 months from data audit to autonomous operation on a first unit; licence plus engineering services; subsequent units cheaper.
• Path B/C/D: 12-24 months and a dedicated team; cost drivers are data cleaning, simulator calibration, OT security, validation and management of change.
• KPIs: margin per tonne, specific energy (GJ/t, kWh/t clinker), yield, off-spec rate, standard deviation of key quality variables, percentage of time AI is in service ("uptime on"), operator override rate, CO2 intensity, number of constraint violations (target zero).

People & change

Begin with plant management and the head of production (process owner) to define economic targets and the allowed action range. Process engineers, control engineers and shift operators are the key end users; involve them from the first data review, because they validate the model and own the operating envelope. IT and IT security must approve the OT/IT architecture (IEC 62443 zones, NIS2); data owners release historian and LIMS data; the safety officer confirms that safety instrumented systems stay independent.

Works council: depends. The process data itself is not personal, but operator actions, overrides and acknowledgements are logged; if these can be linked to individuals, §87(1) no. 6 BetrVG applies. Even otherwise, §90(1) no. 3 requires timely information about AI-planned work processes, and §80(3) allows an AI expert. Public-sector sites involve the staff council. GDPR exposure is small, but ask the DPO to confirm that operator logs are pseudonymised or aggregated (GUIDE-005).

AI literacy: support operators with simulator-based training on what the agent does, how to read its advice and how to switch it off (Art. 4 AI Act, GUIDE-006). Communicate that operators keep the last word: shadow mode first, then advisory mode, then supervised closed loop on one unit as pilot group, with bumpless fallback. Typical fears are loss of control and blame for AI errors; address them with clear authority rules and full audit logs. See GUIDE-008, GUIDE-009 and GUIDE-013.

Risks, regulation & compliance

• EU AI Act: classification depends. Process control in refineries, chemicals and cement is not listed per se in Annex III. Annex III point 2 covers AI as a safety component in the management and operation of critical digital infrastructure and of the supply of water, gas, heating and electricity, so AI steering district heating, gas or power assets can be high-risk. AI that is a safety component of a regulated product (e.g. under the Machinery Regulation (EU) 2023/1230) can be high-risk via Annex I. Plain optimisation of non-safety setpoints is generally minimal-risk, but check each case; Art. 4 AI literacy applies [6]. No Art. 5 practice is involved.
• Functional safety: IEC 61511 (process industry) and IEC 61508; the AI must be outside the safety function, with documented independence and proof tests [5].
• Major-hazard rules: Seveso III Directive 2012/18/EU for sites with hazardous substances (changes to control philosophy need safety-report updates and management of change); IED 2.0 permit limits and EU ETS monitoring for emissions.
• Cybersecurity: NIS2 (energy, chemicals and manufacturing of chemicals as covered sectors), EU Cyber Resilience Act for products, IEC 62443 zoning; an AI write path is a high-value attack target.
• GDPR: essentially no personal data; operator action logs are the only point (works council co-determination, e.g. German BetrVG section 87, when performance data could be linked to individuals).
• Mitigations: shadow mode first, hard bounds enforced in the DCS, watchdog and bumpless fallback, change control for model updates, full audit logs.

Getting started

1. Pick one unit with measurable economics, good instrumentation and an operator champion; write down the baseline (energy, yield, quality variance) for at least 3 months.
2. Audit data: tag quality, sampling, lab-latency, frozen sensors; build or refresh a unit simulator.
3. Run in advisory mode for several weeks and compare operator actions with AI recommendations.
4. Define the safety envelope, independence from the SIS, OT-security design and management of change; get HAZOP/LOPA review sign-off.
5. Enable bounded write-back on one loop, run A/B tests against APC, then widen scope.

Sources

1. Yokogawa press release via Business Wire, "In a World First, Yokogawa's Autonomous Control AI Is Officially Adopted for Use at an ENEOS Materials Chemical Plant" (2023) — https://www.businesswire.com/news/home/20230329005001/en/In-a-World-First-Yokogawas-Autonomous-Control-AI-Is-Officially-Adopted-for-Use-at-an-ENEOS-Materials-Chemical-Plant
2. ARC Advisory Group, "Yokogawa's Autonomous Control AI Adopted for Use at an ENEOS Materials Chemical Plant" — https://www.arcweb.com/blog/yokogawas-autonomous-control-ai-adopted-use-eneos-materials-chemical-plant
3. Turbomachinery Magazine, "Yokogawa Electric Deploys AI Solution at Aramco's Fadhili Gas Plant" — https://www.turbomachinerymag.com/view/yokogawa-electric-deploys-ai-solution-at-aramco-s-fadhili-gas-plant
4. Yokogawa, "Enabling Autonomous Operations" — https://www.yokogawa.com/solutions/featured-topics/ia2ia/enabling-autonomous-operations/
5. IEC 61511 functional safety, safety instrumented systems for the process industry — https://webstore.iec.ch/en/publication/5527
6. Regulation (EU) 2024/1689 (AI Act), Annex III — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Directive 2012/18/EU (Seveso III) — https://eur-lex.europa.eu/eli/dir/2012/18/oj
8. Imubit, Closed Loop AI Optimization — https://imubit.com