A modified version of the linear quadratic Gaussian controller is presented, which is built upon the prediction form of the model. This implies that the controller is more capable of handling non-stationarities, like time-varying model parameters, than the classical type of linear quadratic Gaussian controller, which usually is based on the solution of the Diophantine or the Riccati Equation (in the state-space case). The linear quadratic Gaussian controller is used in a simulation study together with a transfer function model, with time-varying parameters, that describes the relations between supply temperature of the water from a district heating plant and the supply temperature at specific locations in the distribution network. The simulation results show that the variance of the differenced control signal can be reduced drastically without affecting the performance of the controller significantly. (au) (11 refs.).