Motivation
Combined-cycle power plants (CCPP) are among the most efficient fossil-fuel power generation technologies, and their integration with hydrogen production via Liquid Organic Hydrogen Carriers (LOHC) represents a pathway to lower-carbon operation. Optimizing such a system is non-trivial:
- Many interacting design variables (turbine parameters, LOHC flow rates, heat integration points)
- Each Aspen Plus simulation takes minutes — too slow for iterative optimization
- Multiple competing objectives (power output, hydrogen production, efficiency) require multi-objective treatment
The solution: build a shared Aspen Plus + Python automation pipeline, then analyse the resulting dataset along two tracks suited to different audiences and goals.
Approach — Two Analysis Tracks
The two tracks share the same process model and DOE dataset; they diverge only in how the dataset is interpreted.
Track A — Linear Regression (Publication track)
The track currently driving the manuscript.
- Aspen Plus DOE dataset → linear regression per output (power output, H₂ production, system efficiency, etc.)
- Coefficient inspection reveals direct sensitivity of each output to design variables (SR, MCH, P, U)
- Penalty-vs-recovery trade-offs read straight from regression slopes per heat-recovery line (Line 4 / 5 / 6 / 9)
- Lower model complexity → cleaner attribution and easier review for journal readers
- Suited to the linearly-dominated operating region observed in Phase 1 grid sweep
Track B — Surrogate Modeling + NSGA-II (Optimization track)
The longer-horizon optimization track, kept in reserve for future scale-up of the study.
- Trained four ML surrogate models (RF, GB, NN, GP) on the same simulation dataset
- Bayesian optimization (single-objective) and NSGA-II (multi-objective) on top of the surrogate
- Pareto-front exploration across efficiency, hydrogen output, and power penalty
- Enables thousands of optimization evaluations per minute once the surrogate is trained
- Reserved for follow-up work where non-linear interactions become dominant
Phase 1 — Grid Sweep (Complete)
4-line heat recovery comparison across the design grid (SR × MCH) per line:
| Line | Location | Temp | Success Rate | Penalty | E1 Impact |
|---|---|---|---|---|---|
| Line 6 | Post-SH gas | *** °C | *** | ~*** kW | None |
| Line 5 | Post-GT gas | *** °C | *** | ~*** kW | None |
| Line 9 | Post-SH steam | *** °C | *** | ~*** kW | None |
| Line 4 | Post-combustion | *** °C | *** | ~*** kW | Linear decrease |
Quantitative values withheld pending publication. Penalty rankings and qualitative findings are summarized below.
Key findings (qualitative)
- Zero-penalty heat recovery zone discovered at Line 6: LOHC integration causes no measurable reduction in CCPP power output
- 4-line penalty ranking: Line 6 < Line 5 < Line 9 < Line 4 (lowest to highest penalty)
- LOHC competes with steam turbine efficiency, not full CCPP efficiency — penalty scope is narrower than initially expected
- LHHW kinetic model validated at base case with physically meaningful heat-limited behavior
- Maximum H₂ output achieved at Line 6 (high MCH, near-maximum SR)
Phase 2 — Linear Regression + Surrogate Comparison (In progress)
Planned dataset shared by both tracks:
- LHS with 4 variables (SR, MCH, P, U), 500–1,000 samples per line
- Track A: per-output linear regression; coefficient and partial-effect interpretation
- Track B: ANN / RF / GPR surrogate comparison; NSGA-II 4-line independent Pareto fronts
- Overlay comparison with base case (no LOHC integration)
Engineering Implications
TBD
Limitations
TBD
→ Built with: Aspen Automation Framework