Load-cycling operation of thermal power plants leads to changes in operating point right across the whole operating range. This results in non-linear variations in most of the plant variables. This paper investigates methods to account for non-linearities without resorting to on-line parameter estimation as done in self-tuning control. A constrained multivariable long range predictive controller (LRPC), based on generalised predictive control (GPC) algorithm, is designed and implemented in a simulation of 200 MW oil-fired drum-boiler thermal power plant. In order to take into account system non-linearity, the LRPC is evaluated using two types of predictive models: approximate single global linear models and local model networks (LMN). As a simpler alternative, single global linear ARIX models were identified off-line with data generated by running the plant simulation over a load profile covering the entire operating range along with suitable PRBS signals superimposed on controls. For more accurate long-range prediction, networks of dynamic local linear models, identified after dividing the whole operating region into a number of zones, were created. The control strategy gives impressive results, when used in controlling main steam temperature and pressure and reheat steam temperature during large rate of load changes light across the operating range. The improvements are apparent in both constant-steam-pressure as well as variable-steam-pressure modes of plant operation. The results obtained with LMNs based LRPC compare favourably to the those obtained with global model based LRPC.