Data-Driven Predictive Control, representing thebuilding as a cyber-physical system, shows promisingpotential in harnessing energy flexibility f o r demandside management, where the efforts in developing aphysics-based model can be significant. Here, predictivecontrol using random forests is applied in a casestudy closed-loop simulation of a large office buildingwith multiple energy flexibility s o urces, therebytesting the suitability of the technique for such buildings.Further, consideration is given to the featureselection and feature engineering process. The resultsshow that the data-driven predictive control, under adynamic grid signal, is capable of minimising energyconsumption or energy cost.