Neural network models of Cryptosporidium parvum inactivation by chlorine dioxide and ozone

Abstract

Several neural network disinfection models for the inactivation of Cryptosporidium parvum by chlorine dioxide and ozone were developed and compared against existing temperature-corrected Chick-Watson models. In this study a back propagation based network-pruning algorithm called structural learning with forgetting has been used to train all neural network models. The neural network disinfection models performed well relative to the competing Chick-Watson models and learned several basic input variable trends established in earlier disinfection studies. Water temperature was found to be a significant process variable and pH less influential for all neural network models. The final disinfectant residual was included as an input variable to incorporate residual decay, and was found to be more relevant for disinfectants with less stable residuals.