Industrial processes generate large quantities of waste, resulting in health problems and adverse environmental impact. In particular, the treatment and reconditioning of wastewater is a complex problem, due to the existence of strong non-linearity effects, time variant parameters and multivariable coupling not allowing the adoption of simple models to predict the process efficiency and the output water quality. In this paper, the ability of Artificial Neural Networks (ANNs) to predict the quality (pH, electrical conductivity, chemical oxygen demand (COD)) of the wastewater coming from a pharmaceutical industry after treatment in a biological plant was verified. Using a commercial ANN software, various network architectures, differing in the
number of hidden layers and nodes, were tested, in order to find an optimised solution in terms of both precision and learning time. The effectiveness of each ANN configuration was verified by the "leave-k-out" method. Even the simplest ANNs tested were able to correctly describe the pH, due to the relative insensitivity of this parameter to the process conditions. Matching the actual variation of the electrical conductivity proved harder, this task being achieved at the expense of a complication in the network architecture. However, the parameter most difficult to reproduce was the COD, which underwent considerable oscillations within the time window considered. The best ANN architecture was made of seven nodes in the input layer, two hidden layers of fifty nodes each, and three nodes in the output layer. By this solution, reasonable predictions were obtained, provided the input parameters were appropriately selected.