This paper describes an improved algorithm for generalised predictive control (GPC) and applies it to a floor radiant heating system in a full-scale outdoor test-room. The floor heating system is first identified by means of recursive least squares techniques associated with a set of supervision rules for robust system identification. The performance of the floor heating system controlled by GPC, on-off and PI controllers is then evaluated through computer simulations, using the identified models. Experimental results show that the room temperatures predicted with the identified models agree well with the measured data. Simulation results demonstrate that the behaviour of GPC is superior to the other two controllers in terms of response speed, minimum offset and on-off cycling frequency.