A physically structured hybrid model predicts chiller performance under conformal uncertainty and protection events

Online optimisation of vapour-compression chillers requires compact models that remain reliable when setpoints, equipment condition, sensor availability and protection logic change. This paper develops a physics-structured hybrid quasi-steady surrogate (PSHQ). Learned heads predict compressor power and outlet-water temperatures; evaporator cooling capacity, condenser heat rejection and COP are reconstructed through fixed water-side relations. The reference PSHQ gives a compressor-power MAE of 0.0224 kW and R² of 0.963 on stratified validation. A no-retraining scope audit shows that including non-steady interpolation records increases the five reported MAE s by 2.04–4.62 times. Controller-confirmed short-cycle, anti-freeze and high-pressure event wrappers reduce power MAE by 67–86% relative to the uncorrected steady predictor. An added-constraint extension lowers energy-residual MAE from 0.1069 to 0.0962 kW and gives a five-channel mean MAE ratio of 0.966 (95% CI : 0.947–0.977). On the independent RP-1043 test runs, R² ranges from 0.9813 to 0.9974 across the five outputs. A complete-day Day-8 test gives R² values of 0.8479–0.9364 and split-conformal coverage of 0.8710–0.9220. PSHQ provides a calibrated and physics-informed framework for quasi-steady inference and protection-event handling across the tested operating envelopes.