Source-linked AI summary
A Plantwide Simulation Library for Control-Oriented Design: Analytical and Data-driven Use Cases
Kyle Territo, Luis A. Briceno-Mena, Jose A. Romagnoli
TL;DR
Plantwide control methods are often compared on fixed simulated flowsheets, and study configurations are rarely reported in a reproducible form. plantbench provides hashed, plain-data configurations and tests a damping-ratio-based network ranking across 2016 closed-loop runs. The ratio ranks networks but detects limit cycles less effectively than an early-response classifier, and it is nearly insensitive to a temperature-loop setting that partly determines whether the plant holds an operating point.
Problem
Plantwide control studies use a limited set of fixed flowsheets rather than design spaces, and reproducible runs are needed to test whether design-point indices predict post-upset operation.
Source: §I · p. 1Method
plantbench represents plant design, tuning, and instrumentation as hashed plain-data configurations and evaluates a damping-ratio ranking in 2016 closed-loop runs across networks, tuning, analyzer deadtime, and upsets.
Source: §I · p. 1Results
At matched recall, the early-response forest found 60 of 82 limit cycles with 5 false alarms, while a damping-ratio threshold raised 173 false alarms.
Source: B. Recognizing a limit cycle early · p. 4Takeaways & Limitations
The damping ratio ranks the networks but is a weak limit-cycle detector, and the temperature-loop integral time partly determines whether the plant holds an operating point despite the ratio’s near-insensitivity to it.
Source: §VI · p. 5The results are model-to-model, and the fidelity of the second-order Padé approximation to a true analyzer delay is not established.
Source: §VI · p. 5
Abstract
from arXiv · showhide
Control structures and operability indices for plantwide processes are developed and compared on fixed simulated flowsheets, and the configuration of a study is rarely reported in a form that another group can reproduce. In this contribution, plantbench, a library of plantwide process simulations in which the design, the controller tuning and the instrumentation of a plant are declared as hashed plain-data configurations, is presented. On a heat-integrated reactor--separator--recycle plant with an open-loop unstable reactor, a heuristic plantwide control design prices four heat-exchanger networks (HENs) by the least damping ratio of the linearized closed loop, and 2016 closed-loop runs over networks, tuning, analyzer deadtime and upsets test that ranking. Results show that the damping ratio orders the networks as the runs do with the network that recovers the most heat reaching 0.013 against 0.157 for the selected one and cycling under a throughput increase, but that it is a weak detector of which operation will oscillate. At the recall of a random forest on the first 100 min after an upset, 60 of 82 limit cycles, a damping-ratio threshold raises 173 false alarms against 5. The integral time of the temperature loop, to which the damping ratio is nearly insensitive, decides in part whether the plant holds an operating point.
I. INTRODUCTION
Plantwide control studies use a limited set of fixed simulated flowsheets, while learning-based control mostly omits plantwide couplings. plantbench presents a reproducible simulation library and uses a case study to test whether a design-point damping ratio predicts post-upset operation.
I. INTRODUCTION
- Fixed test problems offer chosen designs and disturbances rather than design spaces, and the absence of standard problems remains an obstacle to routine design-control integration.Whether a design-point operability index predicts behavior after upsets involving valve saturation and integrator windup is an empirical question requiring repeatable runs across a design space.Source: §I · p. 1
- plantbench declares plant designs, loop tuning, and analyzer deadtime as hashed plain-data configurations and expands specifications into reproducible datasets.A case interface checked by a library contract lets users bring their own plants under the same protocol.Source: §I · p. 1
- The demonstration prices four heat-exchanger networks by the least damping ratio, then tests that ranking in 2016 closed-loop runs across tuning, analyzer deadtime, and six upsets.The evaluation compares damping ratio with realized upset outcomes, including limit-cycle detection and early-response classification.Source: §I · p. 1
A. Cases and configurations
A plantbench case defines a plant model, steady-state design point, verified reference solution, and design space. Configurations select and validate a case’s control and instrumentation choices, while the closed loop remains linearizable for analysis.
A. Cases and configurations
- A case combines a plant model, steady-state design point, verified reference solution, and design space.The library applies configurations, datasets, tasks, and sensitivity analysis to cases implementing its interface.Source: A. Cases and configurations · p. 1
- A configuration selects a design-space member through nine plain-data sections, round-trips through TOML, and is hashed to identify its run.Each section is checked against the case when the plant is built, so a misspelled loop or parameter raises an error.Source: A. Cases and configurations · p. 1
- PI loops and ratio stations define control structures, while measurement and valve non-idealities preserve a closed loop that can be linearized.The least damping ratio is computed over modes with |Im λ| > 0.05 min^-1; controller integrals are appended to the plant state and windup uses back-calculation.Source: A. Cases and configurations · p. 2
B. Specifications, tasks and verification
Dataset specifications and tasks make a study’s sampled configurations, evaluation protocol, and run identity reproducible. The paper reports its dataset in that form.
B. Specifications, tasks and verification
- A dataset specification names a case, seed, and base configuration, combines partial configurations as a Cartesian product, and draws a Latin hypercube over continuous paths.The expansion is deterministic in the seed and its manifest records the specification, library version and commit, and source-tree cleanliness.Source: B. Specifications, tasks and… · p. 2
- A task fixes the signals and window available to a method, its labeling rule, folds, and metrics, so methods with the same protocol digest use the same folds.The protocol digest combines the specification and task; the run digest uses the specification alone.Source: B. Specifications, tasks and… · p. 2
- Table I reports this paper’s dataset in the library’s specification format.The case reproduces reactor steady-state and eigenvalue reference values and byte-for-byte reference results under the 329-test library suite.Source: B. Specifications, tasks and… · p. 2, Table I · p. 2
III. CASE STUDY AND DESIGN PROBLEM
The case study is a heat-integrated reactor–separator–recycle plant with an open-loop unstable reactor and four candidate heat-exchanger networks. Its design problem balances profit and energy-related environmental impact against reliable operation under specified upsets.
III. CASE STUDY AND DESIGN PROBLEM
- The 134-state plant uses an open-loop unstable CSTR and a thirty-stage separation column with recycle and purge.Its reference structure fixes reactor effluent flow, controls reactor level on fresh feed and tray 17 temperature on boil-up, and adjusts reactor-temperature setpoint through a composition loop.Source: §III · p. 2
- Four HENs offer different heat-recovery arrangements, while the design problem seeks high operating profit and low energy GWP with reliability under six specified upsets.The choices include D4’s pump-around of reactor heat to the reboiler and variations in reactor-temperature-loop tuning and analyzer deadtime.Source: §III · p. 2
- Runs are classified from reactor temperature over the final 300 min as limit-cycle, unsettled, saturated, or settled.A limit cycle requires a swing above 0.5 K peak-to-peak with at least half of that swing sustained across the final two 150-min windows.Source: §III · pp. 2–3
IV. ANALYTICAL PLANTWIDE DESIGN
The analytical design procedure compares control structures and heat-exchanger networks using damping and upset response. It selects D3, while D4’s high heat recovery comes with low damping and cycling under a throughput increase.
IV. ANALYTICAL PLANTWIDE DESIGN
- With a 10% fresh-feed increase, recycle growth falls from 5.22 times the production increase with no flow fixed to 1.46 times once the composition cascade holds conversion per pass.Fixing the reactor effluent flow alone yields 4.51 times the production increase.Source: §IV · p. 3
- At Kc = −2 m3/(min K) and τI = 5 min, the least damping ratio is 0.061; at Kc = −1 and τI = 20 min, it is 0.158.The reactor temperature loop is tuned for damping rather than stability alone.Source: §IV · p. 3
- D4 recovers the most heat and reaches the lowest energy GWP, but its least damping ratio is 0.013 against 0.155–0.157 for the other designs.All four designs share a rightmost eigenvalue of −0.0028 min−1; D4’s pump-around leaves the jacket at 0.267 m3/min against 1.092 m3/min in the other designs.Source: §IV · p. 3
- Under a 10% throughput increase, D4’s reactor-composition integral absolute error is 291.24 kmol min/m3, versus 20.20 kmol min/m3 for D1, D3, and D5.D4’s jacket valve is limited for 75% of the run, and its reactor temperature oscillates between 356.05 and 371.87 K; the other integrated designs settle.Source: §IV · p. 3, Figure 2 · p. 4
- D3 is selected because it meets the minimum-utility target of pinch analysis at its pressure with one exchanger and outperforms D5 on both objectives.The procedure leaves D4 at a high pump-around fraction as a design known to cycle.Source: §IV · p. 3
A. The enumerated dataset
The enumerated dataset spans multiple plants, upsets, tunings, and analyzer deadtimes; most runs settle, while limit cycles are concentrated in D4 and are only partly distinguished by damping ratio.
A. The enumerated dataset
- Of 2016 runs across seven plants, six upsets, and 48 tuning and analyzer-deadtime settings, 78.4% settled, 17.5% remained unsettled, and 4.1% ended in a limit cycle.All runs completed in 6.89 h of integration, and none ended saturated.Source: A. The enumerated dataset · p. 4
- Limit cycles concentrate in D4: 3 of 288 runs cycled at pump-around fraction 0, versus 53 of 288 at 0.7; 50 of 82 cycles followed the 20% throughput increase.Unsettled runs were almost entirely associated with the inert upset, which moves the plant on a timescale longer than the run.Source: A. The enumerated dataset · p. 4
- The damping ratio has ROC AUC 0.872 for detecting limit cycles, but cycles occur up to ζ = 0.281 while 1803 runs at or below that value do not cycle.Median damping ratio is 0.036 among limit cycles versus 0.166 among settled runs; median integration time is 86.60 s for cycles versus 3.52 s for settled runs.Source: A. The enumerated dataset · p. 4
B. Recognizing a limit cycle early
The paper tests whether signals from the first 100 minutes after an upset predict the late outcome across held-out plant–upset combinations. Early-response random forests outperform design-point indicators on that grouped evaluation.
B. Recognizing a limit cycle early
- Each run contributes 176 features from 16 signals sampled every 10 minutes as means over 8 minutes during the first 100 minutes after the upset.The evaluation compares 10-NN and a 500-tree random forest with design-point damping ratio, rightmost eigenvalue, and design-variable baselines on stratified and plant–upset-grouped folds.Source: B. Recognizing a limit cycle early · p. 4
- On grouped folds, the early-window forest reaches balanced accuracy 0.884 and finds 60 of 82 limit cycles with 5 false alarms.The folds hold out whole plant–upset combinations so test-combination runs are absent from training.Source: B. Recognizing a limit cycle early · p. 4
- To find 60 limit cycles, a damping-ratio threshold of 0.0744 raises 173 false alarms, while the forest using design variables finds 14 limit cycles on grouped folds.At matched recall of 28, the damping-ratio threshold raises 17 false alarms versus 0 for 10-NN; the threshold is read on all runs, favoring damping ratio.Source: B. Recognizing a limit cycle early · p. 4
C. Mechanism and sensitivity
A data-mining workflow recovers operating paths, while sensitivity analysis shows that the damping ratio and realized operation respond to different design variables.
C. Mechanism and sensitivity
- The workflow embeds the last 300 min, clusters runs, then searches for separating variables without using design variables or outcomes as inputs.It recovers the two paths identified in Section IV.Source: C. Mechanism and sensitivity · p. 4
- The sampled design space spans pump-around fraction [0, 0.7], column pressure [0.15, 0.275] bar, Kc ∈ [−4, −0.25] m3/(min K), τI ∈ [5, 160] min, and analyzer deadtime [0, 10] min.Source: C. Mechanism and sensitivity · p. 5
- 25.9% of cluster-separating contributions came through reactor inlet temperature; using run-centered deviations, the largest unsettled cluster differed through jacket coolant flow at 67.7%.The first comparison uses deviations from the design point; the second uses deviations from each run’s own late-window mean.Source: C. Mechanism and sensitivity · p. 5
- 0.203 against 0.736 for deadtime’s first-order and total indices shows interaction effects; integral time’s total index is 0.025 for damping ratio and 0.305 for outcome.Across D4 designs at a 10% throughput increase, damping ratio responds most to temperature-loop gain and analyzer deadtime, while operation also responds to pump-around fraction and integral time.Source: C. Mechanism and sensitivity · p. 5, Table IV · p. 5
VI. CONCLUSIONS
The library reports hashed plain-data study configurations; on the heat-integrated case, damping ratio ranks network outcomes but poorly detects cycles. Operating-point retention also depends on temperature-loop integral time, nearly invisible to the design-point damping ratio.
VI. CONCLUSIONS
- 0.013 against 0.157: the damping-ratio ranking matched 2016 closed-loop runs in rejecting the network that recovered the most heat.A random forest on the first 100 min after an upset found 60 of 82 limit cycles with 5 false alarms; a damping-ratio threshold at the same recall raised 173.Source: §VI · p. 5
- Hashed plain-data configurations identify the library’s plant designs, tuning, instrumentation, runs, and study protocol.Source: §VI · p. 5
- Results are model-to-model with simulator-defined outcomes; the second-order Padé deadtime approximation has unestablished true-delay fidelity, and inert-upset dynamics exceed the 1500 min horizon.Source: §VI · p. 5
- Operating-point retention depends on temperature-loop integral time, to which the design-point damping ratio is nearly insensitive.Source: §VI · p. 5