AutoModelling: From BMS operational data to validated Modelica models for chillers

Generating executable chiller models from raw Building Management System (BMS) data remains difficult in practice because operational datasets are heterogeneous, incomplete, noisy, and rarely organised for model identification. This study develops AutoModelling, a standardised and reproducible automated workflow that transforms raw BMS operational data into validated, executable Modelica chiller models at the device level within a single pipeline, integrating semantic point mapping, operational data preprocessing, moving-window steady-state screening, EIR-family model identification, rule-based Modelica file generation, automated simulation, and quantitative validation. Rather than imposing one predefined structure, the workflow applies a candidate-model strategy that selects the better-supported variant per chiller from two Energy Input Ratio (EIR) family models in the Modelica Buildings Library—ElectricReformulatedEIR (EEIR) and ElectricEIR (EIR)—which differ in whether the condenser-side performance curves use the leaving or entering condenser-water temperature. The workflow is evaluated on three real cooling plants in Hong Kong, representing a data centre, a commercial complex, and a university campus. Among 22 chillers, 15 provide sufficient complete and steady-state data for identification; EEIR is selected for 11 units and EIR for 4 units. The selected models satisfy the adopted ASHRAE Guideline 14 thresholds for compressor power in 13 of the 15 retained chillers and for coefficient of performance in 10, with 9 chillers meeting both criteria simultaneously. A central finding is that modelling readiness is governed not by nominal sensor availability alone but by key-variable completeness, retained steady-state sample count, operating-range coverage, and measurement consistency: the cleanest site achieves compliant fits from informative data, whereas the others reveal how limited operating coverage or measurement anomalies degrade validation quality. Relative to a conservative manual baseline, the workflow reduces modelling effort by 98.5% for the most demanding site, while transparently reporting both compliant and non-compliant retained units.