Addressing the deficiencies of data-driven heating, ventilation, and air conditioning (HVAC) models in multi-objective joint prediction, long-horizon prediction accuracy, and uncertainty quantification, this study proposes a heterogeneous deep ensemble framework. Through co-simulation technology, a multi-operating-condition operational dataset for a typical variable air volume (VAV)–Reheat system was constructed. By adopting an operating-condition-level partitioning strategy, the input features were strictly decoupled into historical system states, internal loads, and external control commands. On this basis, the framework integrates the advantages of Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), and Temporal Convolutional Networks (TCN). Utilizing an inverse-variance weighting mechanism, it achieves the synchronous and accurate prediction of 11-dimensional targets, encompassing total system power, five-zone temperatures, and five-zone CO₂ concentrations. Simultaneously, the prediction divergence among the ensemble models is utilized to quantify epistemic uncertainty, and a variance scanning mechanism is introduced to calibrate the prediction confidence intervals. The results indicate that within a 6 h prediction horizon, the absolute errors of the model in power, temperature, and CO₂ are reduced by 80.17%, 63–72%, and 75–80.62%, respectively, compared to the physical baseline. On the independent test set, the actual coverage rate of the calibrated prediction intervals approaches 90%, exhibiting an adaptive enveloping variation characterized by tightening during steady states and widening during transients. The proposed heterogeneous ensemble framework demonstrates superior prediction accuracy and long-horizon prediction stability, providing reliable model support for multi-objective optimal operation and robust control of buildings.