Application of AI and deep reinforcement learning (DRL) to process optimisation across refining, blending, and hydrogen systems
Application Controlled
Manipulated variables (MVs)
Disturbance variables (DVs)
AI/DRL value contribution Learns 'golden
variables (CVs)
CDU/VDU
• Product cut quality (IBP, 10%, 90%, FBP) • Kerosene flash point
• Furnace duty (COT)
• Crude slate variability • Salt and water slugs • API, TAN changes
• Reflux and
setpoints' balancing yield, energy, and corrosion risk. Anticipates crude changes and adjusts operation before quality drifts, reducing energy usage and increasing throughput safely.
pumparound rates
• Diesel density
• Cut point temperatures • Heat exchanger fouling
and sulphur
• Stripping steam • Crude throughput
• Ambient conditions
• Salt and water after
desalter
• Column top/bottom
temperatures
• Energy consumption
Gasoline blending
• RON/MON
• Component flow
• Tank stratification
Learns optimal blend
• RVP
ratios
• Component quality drift recipes that minimise
• Density
• Trim stream usage (butane, reformate, alkylate, ethanol) • Header routing
• Inventory constraints • Market spec changes
giveaway while guaranteeing
• Sulphur and emissions
specs
compliance.
• Final blend compliance
Continuously adapts to tank variability and drives blending closer to economic optimum. Learns safe and efficient operating envelopes. Maintains purity with minimum purge loss, detects abnormal crossover early, and protects stack lifetime while maximising hydrogen efficiency.
Hydrogen/
• O₂ in H₂ (or pO₂)
• Purge rates
• Load changes
electrolysers • H₂ in O₂ (crossover)
• Dryer operation • Operating envelope
• Pressure fluctuations • Membrane aging • Water management
• Purity specifications
• Safety margins • Stack efficiency
limits
• Start-up/shutdown
variations
indicators
sequences
Table 2
compliance. In hydrogen systems, they maintain purity and safety margins with minimum purge losses and maximum efficiency. This fundamentally changes the role of analysers. Measurement becomes experience, experience becomes intelligence, and intelligence becomes continuously improving operation (see Table 2 ). economics of process analytics. By relocating sensitive electronics to control rooms and using in-situ or fibre-optic sensing, they dramatically reduce capital investment. Analyser shelters, hazardous area enclosures, long sample lines, purging systems, and complex conditioning Capex, Opex, and BOP optimisation Smart optical analysers also redefine the
hardware can often be eliminated or simplified. A single optical platform can serve multiple measurement points, reduce duplication, and simplify project execution. The impact on Opex is even greater. Mechanical sampling systems are maintenance-intensive assets. Their reduction improves analyser availability, lowers calibration effort, and reduces unplanned downtime. Beyond the core process units, the balance of plant (BOP) is a major optimisation space. Pumps, compressors, dryers, heat exchangers, purge systems, cooling loops, and utilities account for much energy use and operational stability, yet are often run conservatively due to limited visibility. With optical analysers and AI, the BOP becomes measurable and controllable. Purge
Refining India
20
Powered by FlippingBook