The diversity of occupant air-conditioning (AC) behavior in buildings leads to significant discrepancies in building energy consumption. Thus, establishing models that reflect diverse realistic occupant AC behavior is crucial for the building performance simulations (BPS). Owing to the lack of large-scale samples, it is difficult to extract typical AC behavior patterns from existing studies. With advances in IT methods for data collection, large-scale on-site air-conditioning (AC) operation data of variable refrigerant flow (VRF) systems have become available. Therefore, based on the Big Data of 1200 indoor units (IDUs) of VRF systems, a novel method was proposed herein to establish realistic AC behavior models and extract typical patterns. The overall framework included three steps: data collection and preprocessing, large-scale occupant AC behavior model establishment, and typical behavior pattern extraction and evaluation. First, data preprocessing approaches were implemented to address data issues. Subsequently, based on Big Data, Bayesian parameter calibration was proposed to construct large-scale occupant AC behavior models. Afterwards, using the large-scale calibrated AC behavior models, the cooling demands of the population were simulated, and k-means clustering analysis was adopted to categorize the AC behavior into five groups representing different cooling demand levels and a typical model reflecting the group feature was selected. In the model validation, a hypothesis-based test was applied to confirm the validity of extracted typical patterns in reflecting the real distribution of district cooling demands. In the results, this study identified two occupancy patterns, i.e. “Daily occupied mode” (DOM) and “Night occupied mode” (NOM), and based on two occupancy types, five typical AC behavior patterns were extracted separately. From the typical patterns, it could be noticed that 90% of occupants had relatively low cooling demand less than 40 kWh/m2 during the cooling season. Moreover, comparing to the performance of fixed AC schedule model in district cooling demand evaluation, the extracted typical AC behavior patterns improved the accuracy up to 30.4% when calculating the bias between average modelled cooling demand and measured data. Additionally, the typical AC behavior models put forward in this study can be further integrated in BPS tools, which is promising for the simulation of cooling demand at building or district scale, and realizing AC technology evaluation in future engineering projects.