This paper presents the results from two real-world trials of an optimised supervisory model predictive control (MPC) system for heating, ventilation and air conditioning (HVAC) in commercial buildings. The system learns a model from historical data and uses weather forecasts and a given temperature set-point profile to predict building zone conditions and thermal comfort with the aim of optimising building controls for a number of HVAC zones throughout a day. The multi-objective optimisation minimises running cost and CO2 emissions, subject to operator preferences, while constraining occupant thermal discomfort to an acceptable range to find the best zone temperature set-point schedule for the building. This schedule is then applied to the building by a feedback control loop, which balances the power supplied to each zone for heating, cooling and ventilation. A complementary online occupant comfort feedback tool was deployed to all occupant computers in the trial office buildings. This tool allows occupants to submit feedback on their thermal comfort and satisfaction at any point in time via electronic surveys, as well as allowing these surveys to be issued to occupants at scheduled times. This feedback fine-tuned the thermal comfort model used to constrain the optimisation, allowing for errors in the comfort model to be compensated. Thermal comfort feedback was also used to measure and compare relative occupant comfort levels with a baseline. This control system was trialled on two office buildings in Australia, over two winter months and results compared with the performance of the incumbent building management and control system (BMCS). An average energy reduction of 19% and 32% was achieved in the two buildings over 51 and 10 days of operation respectively without substantially affecting measured or modelled occupant thermal satisfaction levels. (C) 2014 Published by Elsevier B.V.