This paper presents the development of three dynamic models of a multi-effect, falling-film evaporator and compares their performance. The models developed are: an analytically-derived model, an artificial neural network and a linear regression model with an ARX (Auto-Regressive with eXogenous inputs) structure. The development of the analytical model follows a systems approach to analysing the process. The paper focuses on the development of the neural network, in particular developing techniques to improve model flexibility and the use of prior knowledge. The neural network was formed by combining submodels, each modelling a specific element of the overall system, resulting in a modular-structured model. The elements to be modelled were selected using prior knowledge of the system. The linear ARX model was structured in a similar manner. It was found that the empirical models had a superior predictive performance over the analytical model. The modular models also provide benefits in terms of model development effort, flexibility and simple implementation within model-based control strategies.