Control room · hypothesis, not an interview

Hold the turbine at 1473 K. Send a fuel setting only when the tests allow it.

This screen is for a control-room operator. That person has not been interviewed. The job assumed here is the one they do today: change the fuel flow when inlet air, solar heat, or the propane fraction moves, so the turbine inlet stays at 1473 K. The numbers come from the model in the course report. Moving a slider does not retrain it.

Reference setting

303 K · propane 0.25 · 1,500 kW

Inlet air temperature 303 K Propane fraction 0.25 Solar heat 1,500 kW

Air–fuel ratio

— —

Efficiency

— % —

A higher air–fuel ratio means less fuel injected for the same air. Typical efficiency error on the tests is 0.002, about 0.2 percentage points. That is an order of magnitude, not a guarantee at every point.

Measured air–fuel ratio against this model’s prediction. The marked point is 318 K, propane 0.5, 2500 kW: 131 measured, about 116 predicted.

How the small network was checked

Eight unnamed measurements, trained two ways, plus the same starts raised by 20%. The learning-rate slider recomputes only this eight-row fit. It does not touch the turbine model.

Newton learning rate 0.015

Saved result. Move the learning rate to recompute this fit.

Weight	Value

Predicted output against the eight measurements.

Training loss of the report model, falling below 0.04

Training trace for the turbine model in the report. The learning rate steps from 0.005 to 0.0001. The normalized error ends at 0.017. This curve is not an operating result.