This paper presents a model that uses non-fully connected Feedforward Artificial Neural Networks (FANNs) for the forecasting of a seawater-refrigerated power plant condenser performance using the heat transfer rate (Q) over dot, the heat transfer coefficient (U) and the cleanliness factor (F-c). The model developed includes FANNs that take into account the previous temporal values of the most important variables for obtaining the condenser performance, in order to forecast the next temporal value, as well as FANNs that relate the forecasted values with the corresponding condenser performance values (Q) over dot, U and F-c. In FANN architectures, the physical relationships between variables were taken into account. To analyze the model's performance, different ways of grouping data were used: high tide and low tide, right side water box and left side water box of the condenser and time step (daily and every three days). The errors in the test stage for (Q) over dot, u and F-c were acceptable, being less than 0.5% for (Q) over dot, around 4% for U and around 2% for F-c. The errors in the forecasting stage for U and F-c increased with respect to the test stage. (C) 2001 Elsevier Science Ltd. All rights reserved.