Universally consistent hybrid regression model for water quality prediction

Abstract

This work is motivated by a specific problem of a modern paper manufacturing industry, in which boiler inlet water quality is a major concern. Boiler helps to produce power and steam which is used to cook wood chips (along with chemicals) to be supplied to a paper machine. If water treatment plant can't produce water of desired quality as specified by the boiler, then it results in poor health of the boiler water tube and consequently affects the quality of the paper. Variation in inlet water quality of the boiler is due to several crucial process parameters. We formulate this problem into a typical regression problem that involves finding crucial process parameters from the set of available suspected causal variables. Then we build a novel nonparametric hybrid model to solve this regression problem and call it as optimal neural regression tree (NRT) model with the aim to solve the issue of water quality. The major advantage of the proposed optimal NRT model is that it has very less tuning parameters and is easily interpretable as compared to "black-box-like" complex machine learning models. Statistical properties including optimal values of the important model parameters of the proposed model are proved in this paper. We have also experimentally assessed the performance of the proposed model in comparison with the other state-of-the-art models.