A regime-aware digital twin for uncertainty-calibrated prediction and replay-based virtual commissioning of variable-capacity vapour-compression systems

Variable-capacity vapour-compression systems expose strong dynamic coupling among compressor speed, valve opening, pressures, heat-transfer boundaries and safety variables. For control use, a one-step fit is not enough. The predictor also needs to track regimes, quantify uncertainty, adapt to drift and screen candidate actions. A regime-aware dynamic twin (RADT) pipeline is developed here from campaign-scale experiments on a variable-capacity vapour-compression rig. Physics-weighted K-means separates the operating regimes, and the 11-output RADT is trained with grouped state variables, pressure-derived features, a condenser-informed feature, ensemble aggregation and physics-corrected fine-tuning. Ensemble intervals are then calibrated by regime, a lightweight residual layer updates key outputs online, and the RADT is embedded in RADT-based real-time iteration sequential quadratic programming (RTI-SQP) for virtual commissioning. The chronological test set contains 1864 transitions from eight campaign blocks. One-step coefficient of determination (R2) reached 0.9855 for compressor power and 0.9125 for evaporator cooling capacity. For suction superheat and discharge temperature, R2 reached 0.9879 and 0.9764. Regime-conditioned conformal calibration increased mean 90% coverage from 83.6% to 96.4%, including a rise from 36.5% to 98.0% for condensing saturation temperature. On a 482-transition late-period shift subset, online correction reduced the mean root mean squared error (RMSE) of the four corrected outputs by 46.5%. In eight-block replay, predicted energy use fell by 3.8%, from 4768.2 to 4587.0 kW-steps. Mean coefficient of performance (COP) rose from 3.318 to 3.455, and superheat violations dropped from 297 to 78. RADT can be used to screen candidate control actions on historical campaign data before plant testing.