Reinforcement‑learning blend trim for DSN column stability and margin.
A Deep Q‑Network that coordinates the distillate blend valves with DSN column level, pulling more high‑margin naphtha into jet and ULSD while holding the tower steady. Loss‑of‑level upsets stop, giveaway turns back into product, and it runs closed‑loop on the plant's existing Yokogawa DCS with no new instrumentation.
A refinery blends DSN column bottoms naphtha into finished distillate, kerosene/jet, high‑ and low‑flash ULSD, where it is worth far more than as reformer charge. But the blend valves and the column's bottoms‑level controller had no coordination: when the blend valves called for more naphtha than the column could give up, the level dropped out, a pressure interlock slammed shut, and the whole blending system swung, roughly once a week, with hours of product giveaway each time.
Ai‑OPs deployed a Koios Deep Q‑Network (DQN) that learns to balance the DSN column against the blending system. It trims the blend valves and level valve together, maximizing naphtha pulled into high‑margin product while keeping the tower in a safe band, and hands the loop hard low‑level guardrails. Low‑level dropouts fell from a weekly event to zero, recovering an estimated 50–75 BPD of incremental distillate worth over $1M per year.
Ai-OPs
·DSN Blending RL Control01
Background & objective
The objective was two‑fold: minimize how often the DSN column bottoms level drops out, and free up the naphtha those upsets, and a conservative minimum‑flow setting, were keeping out of high‑margin blending.
DSN column bottoms is a light naphtha that can go two ways. The level valve sends it to reformer charge, the low‑value disposition that also protects column level. The blend valves pull it into finished distillate, kerosene/jet, high‑ and low‑flash ULSD, on ratio setpoints, where every incremental barrel is worth far more. The two compete for the same naphtha, and nothing coordinated them.
The control problem
- No coordination: the blend valves chase ratio setpoints with no awareness of column level; the level controller defends level with no awareness of blend demand.
- Weekly dropouts: blend valves over‑pull, the level valve hits its limit, the tower drops below its floor, and a pressure interlock slams the blend header shut to restore level.
- Hours of giveaway: each interlock swings the whole downstream system, resets blend ratios, and gives away margin until it settles.
- Idle capacity: an always‑on minimum‑flow soft‑stop spilled naphtha to reformer charge, pump capacity that was worth more in the blend lines.
Stability alone protects level by blending less. The win is the opposite: blend right up to the column's limit without ever crossing it, which means trading off level and blend flow continuously, in real time.
FIG. 1: BOTTOMS DISPOSITION. The same naphtha can leave as reformer charge through the level valve LV or as finished distillate through the blend valves on the pressure‑controlled (PC) header. Blending wins on margin, but only if the column level holds.
Ai-OPs
·DSN Blending RL Control02
Closed-loop control architecture
A single Deep Q‑Network learns the trade‑off the operators never could hold by hand: it watches column level, blend flow and header pressure, and trims the blend and level valves together, pushing blend flow to the edge of what the column can supply, and no further.
FIG. 2: CONTROL & CONNECTIVITY ARCHITECTURE. One DQN drives three coordinated moves, blend trim, level valve, and a hard low‑level override. It reaches the field through a high‑availability OPC path: the Yokogawa Centum VP DCS, a redundant pair of Yokogawa ExaOPC servers, and the on‑prem Koios edge, both servers linked for failover.
What the agent watches
Column bottoms level, DSN‑to‑distillate flow, blend‑header pressure and the ratio demand on each blend stream. Koios manages the historical data, data quality and data‑range tolerances; the DQN then inferences once per second on the prior 10 minutes of runtime, so it acts on a trend rather than a single noisy sample.
What it controls
It trims the three blend flow controllers and the level valve in concert, and respects a low‑level override that caps blend flow whenever the column nears its floor. The result is a controller that blends to the system's true limit instead of a conservative margin away from it.
Ai-OPs
·DSN Blending RL Control2.3
What the controller balances
The DQN has to balance two objectives that pull in opposite directions, exactly the tension operators had to referee by hand. Holding that balance is the whole job.
FIG. 3: CONTROL OBJECTIVES. The agent earns margin only when the column is safe, so it learns to ride the column's real limit instead of a conservative setpoint, and the hard override means a bad state is never reachable.
03
Implementation & integration
No new hardware
The coordination the project had originally scoped as extra instrumentation was delivered entirely in software. The DQN reads existing tags and writes to the existing blend, level and override loops, so the modification reached the field with no new instrumentation and no capital build‑out.
Deploy & run
The model was trained and validated offline against historical upsets, then promoted from monitoring to closed‑loop control. The control scheme was commissioned mid‑2023; within three months the loss‑of‑level events that had defined the unit were gone.
Operators enable or disable the agent per loop from the standard faceplate, and the low‑level override sits underneath it at all times, full transparency, with the column's safety never delegated to the model alone.
Ai-OPs
·DSN Blending RL Control04
Results
Once the agent took the loops, the weekly loss‑of‑level events that had capped blending collapsed, to four in the first month live, then one, then none, and the recovered naphtha showed up as margin.
FIG. 4: LOSS-OF-LEVEL FREQUENCY. Recreated from the plant event log. Before commissioning the tower dropped out several times a month, with a 61‑event spike during one unstable period. After the DQN took the loops, frequency fell to 4, then 1, then 0, the upset mode that defined the unit effectively closed out.
"It eliminated a $300k/yr margin loss from loss‑of‑level events. By improving blending stability it not only decreases upset downside, but increases steady‑state upside by making it easier to blend fully, close to the limits of the system… Even at a total 75 BPD, that's $1 million a year."
Process optimization engineer, refining operations
Ai-OPs
·DSN Blending RL Control05
Where this applies
The same pattern, a reinforcement‑learning agent balancing competing objectives across coupled loops, closed‑loop on the existing control system, generalizes across refining and petrochemical operations. Applications available today on Koios:
About Koios
Koios is Ai‑OPs' self‑hosted machine‑learning inferencing platform for deploying scalable AI models on industrial control systems. It is hardware‑agnostic, runs fully on‑premises and air‑gapped, and writes optimized control back to PLCs, SCADA and smart instruments in real time. Your models and data stay inside your plant; you own every model Koios trains.
It installs as a single Docker image from a Docker Verified Publisher, or as an offline ISO / virtual‑appliance image, and is licensed as an annual per‑instance subscription.
| Protocols | EtherNet/IP · OPC‑UA · Modbus TCP · BACnet & more |
| Deployment | On‑prem · air‑gapped |
| Install | Single Docker image / ISO |
| Min. hardware | 4‑core 64‑bit · 8 GB RAM · 128 GB SSD · 1 GbE |
| Per license | Unlimited devices, tags & models · historian & diagnostics · support |