During the 1990s a large-scale study of Port Phillip Bay, Australia, was undertaken by the CSIRO (the Commonwealth Scientific and Industrial Research Organisation, Australia's national research body). A major outcome of the study was a complex ecosystem model intended to provide scientific input into management decisions concerning the nutrient load to the bay. However, its development was costly and time-consuming. Given this effort, it is natural to seek smaller models (reduced models) that reproduce the performance measures of the large model (the full model) that are of interest to decision makers. This thesis is concerned with identifying such models. More generally, this thesis is concerned with developing methods for identifying these smaller models. Several methods are developed for this purpose, each simplifying the full model in different ways. In particular, methods are proposed for aggregating state variables, setting state variables to constants, simplifying links in the ecological network, and eliminating rates from the full model. Moreover, the methods can be implemented automatically, so that they are transferable to other ecological modelling situations, and so that the reduced models are obtained objectively. In the case of the Port Phillip Bay model, significant reduction in model complexity is possible even when estimates of all the performance measures are of interest. Thus, this model is unnecessarily complex. Furthermore, the most significant reductions in complexity occur when the methods are combined. With this in mind, a procedure for combining the methods is proposed that can be implemented for any ecological model with a large number of components. Aside from generating reduced models, the process of applying the methods reveals insights into the mechanisms built into the system. Such insights highlight the extent to which the model simplification process can be applied. Given the effectiveness of the model simplification process developed here, it is concluded that this process should be more routinely applied to large ecosystem models. In some cases, the full sequence of methods might prove too computationally expensive to justify its purpose. However, it is shown that even the application of a subset of the methods can yield both simpler models and insight into the structure and behaviour of the system being modelled.