Standard library or domain libraries and literature? A controlled comparison of two Modelica models of the same industrial pump unit on open plant data

A physics-based model of a specific machine can be assembled from the Modelica Standard Library and the equipment documentation, or from domain libraries and published measurements as well. We measured what the second option adds. The object is a boiler feed pump unit: a seven-stage centrifugal pump with a 110 kW induction motor, observed in eight open telemetry campaigns of the CIRA P507 station. Two configurations of our model-synthesis method were compared under criteria registered before computation. They share the object, the data, the acceptance regimes and the calibration and validation protocol: • configuration I uses the standard library and the documents; • configuration II uses eight libraries and a literature base on the thermal behaviour of fan-cooled motors. Model II is built from ThermoPower components with IAPWS-IF97 water. It passed all 148 behavioural acceptance checks without manual changes, on its second build. The literature halves the bias of the prediction from documents alone: median bias 7.8 → 3.3 K, median RMSE 8.1 → 6.3 K. The gain comes mostly from heat transfer of the stopped frame. After calibration on telemetry the two configurations converge: • held-out RMSE is 2.00 against 2.05 K; • heat-soak error is 1.82 against 1.85 K; • identifiability is the same, and the same unexplained factor of two remains. The libraries change the water side, which the plant sensors do not observe. Water heating in the pump is 0.76 K instead of 1.6 K, and it agrees with an IF97 estimate to 0.02 K. The frame temperature changes by less than 0.13 K, and computing time grows fivefold. The registered verdict is “does not improve”. Two findings concern domain components themselves: • they carry assumptions of their own: the pump volume balance omits the pressure work; • they remove a natural guard: a dead-headed pump is computed into superheated steam instead of stopping. Literature priors pay off before calibration and for risk bounds. Domain libraries pay off where the fluid side is the object of monitoring. For calibrated thermal monitoring with one frame sensor the standard library is sufficient.