A fast growing cascade neural network for BOD estimation

In this paper, a fast growing cascade neural network (FGCNN) is proposed, as a software sensor, to rapidly estimate the biochemical oxygen demand (BOD) in wastewater treatment plants (WWTPs). Firstly, a novel method, based on the orthogonal least squares (OLS), is put forward to add input and hidden units to the existing network one by one. Every unit added to the network affords the maximal reduction of the sum of squared errors (SSE). Then, the FGCNN incrementally updates its output weights by iterations without gradients and generalized inverses, while the other weights remain unchanged during the growth of the network. The simple and effective training method make the FGCNN learn extremely fast. Finally, the proposed FGCNN is applied to estimate the BOD in WWTPs using other easy-to-measure or secondary variables. The experiment results show that the FGCNN has better performance on real-time estimation of BOD than other similar methods.