From Constraint to Competitive Advantage: How a European Refinery Transformed Jet Fuel Production with Modcon.AI

Introduction: When Market Signals Outpace Traditional Control

European refineries are operating in an environment where market demand shifts faster than traditional process control can respond. The rapid post-pandemic recovery in aviation, combined with volatile crude slates and tightening product specifications, has exposed a structural limitation in conventional refinery optimisation approaches—particularly in atmospheric crude distillation units (CDUs).

This article presents a real industrial success story from a major European refinery that leveraged advanced AI-driven optimisation from Modcon Systems Ltd. to convert operational complexity into measurable economic value. The results demonstrate not incremental improvement, but a step-change in refinery performance, achieved under full industrial constraints.

The Challenge: A CDU Locked in the Past

The refinery faced an urgent business requirement: increase jet fuel production to capture surging margins, without compromising safety, stability, or downstream units.

However, the atmospheric CDU was effectively locked into historical operating patterns, with multiple structural constraints:

• Jet fuel yield plateaued at 11.8%, well below the 15% target
• Only 77% of jet fuel met full specification
• The unit processed 5–7 different crude types daily, with frequent slate changes
• Operators managed:
  • 59 trays
  • 4 product draws
  • Competing quality constraints: flash point, freeze point, smoke point
• Any sidestream adjustment required 2–4 hours to stabilise
• Manual optimisation was infeasible: Change one variable → three others drift off-spec

Traditional trial-and-error control simply could not cope with:

• High-dimensional interactions
• Non-linear process behaviour
• Real-time price volatility between gasoline, jet fuel, and diesel

The operational burden on the control room was extreme—and increasingly unsustainable.

See also: How the First Home Buyer Grant Helps You Purchase

Why Conventional MPC Reached Its Limits

The refinery had already explored advanced process control and classical Model Predictive Control (MPC) strategies. The limitations became clear:

• Optimisation problem scale exceeded practical MPC capabilities
• Solving a full economic optimisation took ~45 minutes
• CDU dynamics evolve over 2–4 hours, requiring faster decision cycles
• MPC struggled with:
  • Changing crude properties
  • Non-stationary economics
  • Strongly non-linear tray-to-tray interactions

In short, the problem was not controllability—it was dimensionality and adaptability.

The Modcon.AI Approach: Learning the Process, Not Just Controlling It

The refinery deployed Modcon.AI industrial process optimization platform combining:

• Deep Reinforcement Learning (DRL)
• Hybrid first-principles + data-driven process models
• Real-time economic objective functions

Key Deployment Principles

1. Zero-risk learning phase
The AI was trained for 6 weeks in a high-fidelity simulation, equivalent to ~15 years of real operation, before any live control.

2. Full-process scope
The system simultaneously optimised:

• 87 process variables
• 32 manipulated variables
• Across all relevant quality, hydraulic, and economic constraints

3. Non-intuitive strategy discovery

Unlike rule-based or MPC systems, DRL explored and validated operating strategies no human operator would attempt, while respecting all safety and operating envelopes.

Real-Time Performance: Speed Changes Everything

Once deployed, the fundamental difference became evident:

• Decision latency: <200 milliseconds
• Continuous adaptation to:
  • Crude quality changes
  • Feed rate fluctuations
  • Real-time product pricing

This speed advantage allowed the AI to anticipate CDU dynamics, rather than react to them after the fact—something fundamentally impossible with slow optimisation solvers.

Results After Nine Months of Operation

The operational and economic impact was decisive:

Yield and Quality

• Jet fuel yield increased from 11.8% → 15.0% (+3.2 percentage points)
• On-spec jet fuel production improved from 77% → 98%
• Specification violations reduced by 92%

Model and Operator Confidence

• 97% prediction accuracy across all crude types
• Operators enabled automatic mode 78% of the time
• Manual interventions dropped sharply

Financial Impact

• Annual value creation:
  USD 8.2–12.5 million
• Payback period:
  <6 months
• 5-year NPV:
  USD 28–43 million

These figures were achieved without hardware modifications, revamps, or throughput increases—purely through smarter operation.

Continuous Improvement: AI That Gets Better Over Time

A critical differentiator emerged post–go-live.

Five months after commissioning:

• The AI system delivered ~5% additional performance improvement
• Achieved through online learning, adapting to:
  • Seasonal feed variations
  • Long-term equipment behaviour
  • Updated economic priorities

This is not static optimisation. It is living, improving control intelligence.

Strategic Insight: Why This Matters for European Refineries

This case illustrates a broader industry lesson:

• Traditional MPC optimises within known boundaries
• Deep reinforcement learning expands those boundaries

In high-dimensional, economically driven refinery problems—especially where crude flexibility and product swings dominate—learning-based control is no longer experimental.

It is:

• Deployed
• Trusted by operators
• Delivering hard financial results

Conclusion: Industrial AI That Pays Its Way

This is not futuristic AI. It is industrial-grade, safety-aware, economically aligned optimisation, running today in a European refinery and generating measurable profit.

For refiners facing:

• Tight margins
• Volatile markets
• Increasing operational complexity

AI systems that learn, adapt, and optimise in real time are rapidly becoming a strategic necessity rather than an option.

Full technical case study is available on request.