Modeling the Disinfection of Waterborne Bacteria Using Neural Networks

摘要

Neural networks offer an alternative approach to conventional mathematical models for modeling the disinfection of waterborne pathogens. The disinfection process was modeled using two different learning methods: backpropagation and simulated annealing. Simulated annealing is a robust method of optimization capable of escaping local optimums and determining global optimums. Gradient descent, which back-propagation is based on, is a more limited method of optimization that is unable to overcome local optimums. Many neural networks were developed using experimental data to model the disinfection of Escherichia coli and Eberthella typhosa using chlorine and chloramines. The neural network models were developed based on back propagation and simulated annealing and achieved comparable performance results. The models that were trained using simulated annealing required substantially more training time. Sensitivity analysis was used to explore the ability of the neural network models to learn known input variable trends for the disinfection process. Saliency analysis was used to rank the relative importance of each input variable. Each model successfully determined the appropriate input variable relationships. Based on the results of saliency analysis, all of the input variables were determined to be relevant to modeling the disinfection process for the studied combinations of disinfectants and pathogens. The disinfection model based on simulated annealing preformed slightly better relative to the model based on back propagation. Given the practical equivalence of performance results, the model based on back propagation is preferred as it avoids significant model training time.