Closed-Loop Control

The base-case TEP is open-loop unstable: held at constant inputs it trips on high reactor pressure within a few hours. The tep_studio.control package adds a decentralized multiloop PI controller (N. L. Ricker, Decentralized control of the Tennessee Eastman Challenge Process, J. Proc. Cont. 6(4), 1996) that keeps the plant on its operating point.

The controller is separate from the simulator core (principle P2) and consumes only published measurements, never the internal 50-state vector (principle P5).

1. Stabilize the plant

from tep_studio.control import ClosedLoopSimulation

result = ClosedLoopSimulation(control_interval=0.0005, horizon=48.0).run()

print(result.stabilized)                       # True
print(result.terminated, result.truncated)     # False True  (ran the horizon, no shutdown)
print(result.peak["reactor_pressure_max"])     # ~2709 kPa  (trip is 3000)

stabilized is True when the run reaches the horizon without a shutdown. The runner separates termination (an endogenous plant shutdown) from truncation (reaching the horizon), matching the simulator's lifecycle semantics (principle P3).

Control interval

The reference uses a fixed Ts_base = 0.0005 h sample period. Running control_interval = 0.0005 reproduces it most closely. The PI loops use the actual elapsed time in their integral term, so a coarser interval still behaves sensibly (faster, lower fidelity).

2. Control structure

The strategy is encoded declaratively in RICKER_MODE1 (see tep_studio/control/registry.py). Every loop is resolved by schema name, not array index.

Group	Controlled variable	Manipulated / target
Reactor temperature	reactor_temperature	reactor cooling-water valve
Reactor pressure	reactor_pressure	purge ratio r5 (not the recycle valve)
Reactor level	reactor_level	separator-temperature setpoint (cascade)
Separator temperature	separator_temperature	condenser cooling-water valve
Separator / stripper level	separator_level, stripper_level	underflow ratios r6, r7
Production rate	stripper_underflow	production index Fp (scales all feeds)
Feed flows	feed_*_flow	feed valves (setpoint = ratio × Fp)
%G / reactant trims	%G, yA, yAC	feed-ratio trims r1..r4

The reactor-level loop sets the separator-temperature setpoint, which drives the condenser coolant. This cascade regulates reactor inventory.

3. Setpoints and configuration

By default the controller seeds bumpless setpoints from the Mode-1 initial state, so the first action matches the base-case valve positions (no startup bump). Override them explicitly:

import dataclasses as dc
from tep_studio import TennesseeEastmanProcess
from tep_studio.control import RickerMultiLoopController, ClosedLoopSimulation

sim = TennesseeEastmanProcess()
meas0, _ = sim.reset(mode="mode1")
defaults = RickerMultiLoopController(); defaults.reset(meas0)

setpoints = dc.replace(defaults.setpoints, production_rate=28.0)   # demand more throughput
controller = RickerMultiLoopController(setpoints=setpoints, enable_overrides=True)
result = ClosedLoopSimulation(simulator=sim, controller=controller, horizon=24.0).run()

Feature flags:

• enable_composition (default True) — the stable yA/yAC reactant-ratio trims and the %G feedforward.
• enable_pct_g_feedback (default False) — the %G→Eadj feedback loop. Off by default: its feedforward constants were tuned for the original TEP (temex.c), while this package wraps the modified TEP (temexd_mod.c), whose composition dynamics differ. Enabling it without retuning destabilizes the plant.
• enable_overrides (default False) — the high-reactor-pressure override (cuts the production index) and high-reactor-level override (recycle valve).

4. Overrides

The decentralized loops hold pressure well, so the high-pressure override is a last-resort safety net that rarely fires. Under a strong upset it activates and caps pressure below the 3000 kPa trip:

import numpy as np
idv13 = np.zeros(28); idv13[12] = 1.0   # reaction-kinetics drift

controller = RickerMultiLoopController(setpoints=setpoints, enable_overrides=True)
result = ClosedLoopSimulation(simulator=sim, controller=controller, horizon=8.0).run(disturbances=idv13)
fired = any(d.overrides_active.get("high_pressure_to_production") for d in result.diagnostics)

Override parameters

The Mode-1 override thresholds/gains are not in the reference .mdl files (only Mode 3's coolant→recycle override is). The values in RICKER_MODE1 follow Ricker (1996) §4 and are tuned to hold the 3000 kPa limit; retune them for a different plant or operating point.

5. Disturbance rejection

import numpy as np
def idv(i): v = np.zeros(28); v[i - 1] = 1.0; return v

result = ClosedLoopSimulation(horizon=24.0).run(disturbances=idv(1))   # A/C ratio step

IDV(1), IDV(8) and IDV(13) are rejected indefinitely. IDV(6) (total A-feed loss) is among the hardest TEP disturbances. The controller sustains the plant well past the open-loop shutdown time but cannot hold it indefinitely.

A time-varying schedule is also supported:

schedule = lambda t: idv(8) if t >= 1.0 else np.zeros(28)
result = ClosedLoopSimulation(horizon=24.0).run(disturbance_schedule=schedule)

6. Views, metrics, and the experiment record

The run carries auditable artifacts for data-driven studies:

from tep_studio.control import build_experiment_record, online_control_view

result = ClosedLoopSimulation(horizon=24.0).run()

# Regulatory metrics, accumulated at full resolution (not downsampled):
print(result.metrics["iae"]["reactor_pressure"])
print(result.metrics["constraint_violation_steps"])
print(result.metrics["production_rate_mean"])

# Causal online view (only the measurements the controller may read at time k):
view = online_control_view(result.results[-1].measurements)

# Reproducible record (principle P6): source revision, process/config hashes,
# seed, solver settings, setpoints, gains, metrics, and the model-leakage policy.
record = build_experiment_record(result, controller, simulator=sim)
print(record.to_json())

7. Run the example

PYTHONPATH=src python3 src/tep_studio/control/examples/ricker_mode1_closed_loop.py

It runs the closed loop for 24 h, prints the stability outcome and metrics, and emits the experiment record.

8. Paper-style figures

Render the closed-loop figures (PNG/PDF/SVG plus a source CSV each), alongside the base-case validation figures:

PYTHONPATH=src python3 -m tep_studio.control.figures

This writes to src/tep_studio/simulation/validation_outputs/figures/:

• fig_open_vs_closed_pressure — open-loop pressure runaway/shutdown vs the flat closed-loop trace.
• fig_closed_loop_stabilization — 48 h multi-panel trajectory of the controlled variables.
• fig_disturbance_rejection — reactor pressure and level under IDV(1), IDV(8), IDV(13) stepped at 1 h.
• fig_pressure_override — the high-pressure override capping the reactor under a kinetics upset at high throughput.