Predicting Suspended Solids in a Wastewater Treatment Plant by means Neural Networks

We present an analytical study about the prediction of Suspended Solids in the effluent of a Wastewater Treatment Plant by means Neural Networks. The inputs variables are constituted by ten state variables and three operational parameters. We have implemented two kind of architectures depending on if we consider only one output variable or five variables that describe the water quality at the effluent of a WWTP: suspended solids, biochemical organic matter, chemical organic matter, total Kjeldhal nitrogen and total nitrogen. The set of data used was obtained by simulation with the European Co-operation Scientific and Technical Research simulation benchmark. Two known learning techniques have been implemented: a classical adaptative gradient which uses back-propagated gradient information to guide an iterative line search algorithm and a Kalman filter which is used to obtain the best estimate of the weights based on the stream of training data. To try to improve performance and generalization we have implemented two methodologies of variable selection based on genetic algorithms and fuzzy systems. The results show that Neural Networks are a powerful tool for predicting suspended solids even with great turbulence caused for instance by a heavy period of rain because errors do not attaint a significant value. Key–Words: Wastewater Treatment Plant, Suspended Solids, Neural Networks, Adaptative Gradient, Kalman Filter, Genetic Algorithms, Fuzzy Systems