Kohonen self-organising map (KSOM) extracted features for enhancing MLP-ANN prediction models of BOD₅

This paper presents the results of developing a model to predict the concentrations of biological oxygen demand (BOD₅), in the effluent of the primary clarifier of an activated sludge wastewater treatment plant, using other easily measurable water quality parameters. The model is based on the Kohonen self-organising map (KSOM) and multi-layered perception artificial neural networks (MLP-ANN). The KSOM was used to extract the features of the measured data and to deal with the effects of noise and missing values. The best map units of the measurement vectors over the KSOM were used as inputs to the MLP-ANN to reduce the effects of noise and uncertainty in the measurement data, and to replace the missing elements in these measurements. The results of the KSOM-ANN modelling strategy were found to be better than those obtained by the MLP-ANN trained using the raw measurement data.