Elsevier

Journal of Hydrology

Volume 591, December 2020, 125574
Journal of Hydrology

Research papers
Radial basis function artificial neural network (RBF ANN) as well as the hybrid method of RBF ANN and grey relational analysis able to well predict trihalomethanes levels in tap water

https://doi.org/10.1016/j.jhydrol.2020.125574Get rights and content

Highlights

  • Linear/log-linear model was not effective in predicting THMs levels in tap water.

  • RBF ANN shows high performance in predicting THMs levels in tap water.

  • GRA can facilitate generating sound RBF ANN models with the fewer factors.

  • GRA and RBF ANN paved a new way to predict DBPs in actual distribution system.

Abstract

Many models have been developed in previous studies for predicting the formation of disinfection by-products (DBPs) in drinking water. However, most of them were linear or log-linear regression models, and generated based on simulated disinfection of source water or treated water in a laboratory other than real tap water, which shows low application potential in practice. In this study, a radial basis function artificial neural network (RBF ANN) as well as the hybrid method of RBF ANN and grey relational analysis (GRA) was proposed to predict trihalomethanes (THMs) levels in real distribution systems. A total of 64 sets of data including THMs levels (trichloromethane (TCM), bromodichloromethane (BDCM) and total-THMs (T-THMs)) and 8 water quality parameters (temperature, pH, UV absorbance at 254 (UVA254), dissolved organic carbon, bromide, residual free chlorine, nitrite and ammonia) were used to train and verify the proposed model. As compared to linear and log-linear regression models (rp = 0.254–0.659; N25 = 46–78%), RBF ANNs for THMs (TCM, BDCM and T-THMs) prediction consistently show higher regression coefficients (rp = 0.760–0.925) and prediction accuracy (N25 = 92–98%), which indicates the high capability of RBF ANN to learn the complex non-linear relationships involved THMs formation. Further analysis shows that RBF ANNs using fewer water quality parameters based on GRA still make excellent performance in THMs prediction (rp = 0.760–0.946; N25 = 92–98%). This result demonstrates that GRA can be an effective technique to facilitate the generation of sound RBF ANN models with fewer factors.

Introduction

Disinfection is a prerequisite to guarantee drinking water safety. However, disinfectant (e.g. chlorine) can react with organic matter (OM) in water, producing a variety of disinfection by-products (DBPs), including trihalomethanes (THMs), haloacetic acids (HAAs), haloacetonitriles (HANs), etc (Krasner et al., 2006, Hong et al., 2008, Hong et al., 2013, Ding et al., 2013, Hur et al., 2014, Li and Mitch, 2018). Both laboratory studies and epidemiology investigations showed that long term exposure to DBPs could increase the risk of cancer and birth defects (Richardson et al., 2007, Rahman et al., 2014, Beane Freeman et al., 2017, Kaufman et al., 2018, Wu et al., 2019). China’s current regulation set the maximum level of dichloroacetic acid (DCAA), trichloroacetic acid (TCAA), trichloromethane (TCM), bromodichloromethane (BDCM), dibromochloromethane (DBCM) and tribromomethane (TBM) in tap water at 50, 100, 60, 60, 100 and 100 μg/L, respectively; moreover, the sum of the ratio of each THM concentration to its respective guideline value should not exceed 1 (MOH, 2006). In the United States, the maximum levels of 4 THMs (sum of TCM, BDCM, DBCM and TBM) and 5 HAAs (sum of monochloroacetic acid, DCAA, TCAA, monobromoacetic acid and dibromoacetic acid) have been set to 80 and 60 μg/L, respectively (Mian et al., 2018).

DBPs formation is a rather complicated process, which is mainly governed by water quality and disinfection conditions. For instance, OM, which is often quantified by dissolved organic carbon (DOC) or ultraviolet absorbance at 254 nm (UVA254) (Selberg et al., 2011, Hong et al., 2013), can act as organic precursors (Nikolaou and Lekkas, 2001, Hong et al., 2008, Hua et al., 2015, Sun et al., 2018). Increase in disinfectant (e.g. chlorine) dose will greatly improve DBPs formation, and high bromide level will shift the DBPs speciation toward more brominated ones (Hua and Reckhow, 2012, Hong et al., 2016, Zheng et al., 2020). Meanwhile, pH can influence the form of disinfectant (e.g. Cl2) and thus affect DBPs formation (Fukayama et al., 1986, Hong et al., 2017). Nitrite and ammonia may also impact DBPs formation because they are chlorine consumer (Hu et al., 2010, Zhou et al., 2019). In addition, high temperature and long reaction time will improve the formation of stable DBPs (e.g. THMs, HAAs), but may reduce the yields of unstable ones (e.g. HANs) (Xie, 2004). DBPs monitoring is a laborious job, which usually involves the use of toxic solvents, complex pretreatment process and expensive instrument (e.g. Gas Chromatography, Gas Chromatography-Mass Spectrometer, etc.) (Gonzalez-Hernandez et al., 2017, Li and Mitch, 2018, Andersson et al., 2019). Moreover, the measurement of DBPs concentration in drinking water usually takes a long time, making it impossible to timely control operation parameters to control DBPs in practice. Therefore, many predictive DBPs models have been developed by linearly/log-linearly regression of various involved factors (Chowdhury et al., 2009, Hong et al., 2016, Lin et al., 2018, Shahi et al., 2019). These regression models are useful in identifying key factors influencing DBPs formation, and may help to guide decision making in the drinking water treatment. However, these models also have some inadequacies that cannot be ignored:

(1) DBPs predictive models are site-specific to a large extent because of the differences in water quality and treatment methods used by different regions/waterworks. For example, Yoon et al. (2003) tried to predict THMs formation by chlorination of Korean source water using DBPs models from United States, but found that the predicted value was much higher than the measured ones. Therefore, for a specific region with specific water source and and treatment process, it is necessary to establish its own DBPs models. (2) Most available DBPs models were developed through simulated disinfection of source water or treated water (from water works) under carful design in laboratory (Chowdhury et al., 2009, Hong et al., 2016, Lin et al., 2018). However, the real drinking water should undergo water treatment processes, disinfection and pipeline transportation. The composition and distribution of DBPs are quite different from those obtained from simulated disinfection of source water/treated water. (3) Most models contained the parameter of “disinfection time”(Sohn et al., 2004, Chowdhury et al., 2009, Hong et al., 2016). But for real water supply system, it is quite difficult to measure the disinfection time (i.e., how long it takes from the beginning of disinfection in water treatment plant to the end of the user), which must consider the distance, season, water supply data of the sampling points, and etc. (4) The relationship between DBPs formation and varied factors were usually weakly linear/log-linear, which made these models not necessarily effective in predicting DBPs formation (Sadiq and Rodriguez, 2004, Chowdhury et al., 2009, Ding et al., 2013, Lin et al., 2018). In some cases, the prediction accuracy (the proportion that the prediction error < 25%) only reached 58–66% (Lin et al., 2018). Because of the above disadvantages, it is quite difficult for the available regression models to be put into practice. It is essentially necessary to establish DBPs models exclusively for real distribution system in a specific area, and consider the complicated non-linear relationships between various factors and DBPs formation in order to develop effective prediction models.

Artificial neural networks (ANN) are powerful tools to deal with complex interactions, and they are considered as standard non-linear estimators (Singh and Gupta, 2012, Iliyas et al., 2013, Yang et al., 2013). Given the complex non-linear relationships between DBPs formation and various factors, ANNs may provide an attractive alternative to predict DBPs formation. Nevertheless, the studies on prediction of DBPs using ANNs are still very limited. Based on the latest information from Web of Science, there are seven studies on ANN modes for DBPs prediction: five used back propagation (BP) ANN (Milot et al., 2002, Kulkarni and Chellam, 2010, Ye et al., 2011, Singh and Gupta, 2012, Park et al., 2018), another two used autoencoder-neural network and hybrid genetic algorithm-based ANN (Moradi et al., 2017, Peleato et al., 2018). These studies demonstrated that ANNs are effective in the prediction of DBPs. However, only two studies were related to actual distribution system (Ye et al., 2011, Moradi et al., 2017), both of which included the parameter of “residence time/contact time” (i.e., how long it takes from the beginning of disinfection in water treatment plant to the end of the user). Contact time is easy to measure in a laboratory rather than in actual distribution systems, where the distance, season and water supply data of the sampling points need to be considered. Therefore, the available ANN models on actual distribution system are quite difficult to be put into practice. Developing ANN models for DBPs prediction in real distribution system without residence time becomes essential. Moreover, tests of different ANNs models are still worth further exploration. The radial basis function (RBF) neural network refers to a kind of feed forward neural network with excellent performance. RBF network can approximate any non-linear function with arbitrary accuracy, and realize global approximation, without any local minimum problem (Jin and Bai, 2016, Zhao et al., 2019). Also, it has a fast learning speed because of the compact topological structure (Han, 2006, Chen et al., 2019). The distinct advantages give RBF ANNs strong application potential in more and more fields (Lin, 2011, Zarbakhsh and Addeh, 2018, Zhao et al., 2019). However, systematic studies on the application of RBF ANNs in DBPs predictions are few (Lin et al., 2020).

On the other hand, water quality parameters and/or treatment conditions serve as input factors during the development of ANN models for DBPs prediction (Ye et al., 2011, Moradi et al., 2017, Park et al., 2018). However, the questions regarding which factors are important or indispensable, and which factors are not necessary remain unclear. Considering the non-linear relationship between various factors and DBPs formation, it might not be suitable to use the traditional method (e.g. regression analysis) to identify the importance of factors. Given the uncertain relationships between DBPs formation and various factors, DBPs formation can be regard as a grey system. Therefore, the gray relational analysis (GRA) specifically used for grey systems may be a good choice to identify the importance of the factors that affect DBPs formation. The GRA is a quantitative method to estimate situational changes among data sequences based on the grey system theory (Liu and Lin, 2011, Tian et al., 2018). It can reveal the non-linear relationship between multiple variables according to similarities in their geometric proximity and rank their relational degrees (indicator for importance) in descending order (Wong et al., 2006, Yang et al., 2013, Deng, 2019). Therefore, GRA can be an effective supplement during the construction of ANN models for DBPs prediction, facilitating to set up an ideal model with fewer factors, and eventually saving time and cost in water quality monitoring for water treatment plant.

The main objective of this study is to develop prediction models of DBPs formation in 13 water supply systems in 8 counties in Jinhua Region of Zhejiang Province in China. Since THMs have been demonstrated to be the most abundant DBPs presented in drinking water and the primary drivers of cancer risk (Li and Mitch, 2018), the models for TCM, BDCM and total THM (T-THMs) were included. Data set on THMs levels and water quality parameters were separated into a training group and a testing group for prediction by multiple linear/log-linear regression (MLR), RBF ANN and GRA-RBF ANN. Finally, the prediction results were evaluated, and the advantages of the proposed GRA-RBF ANNs over RBF ANN and MLR for DBPs prediction were discussed.

Section snippets

Description of dataset

The dataset used in this study were originated from our previous study (Zhou et al., 2019). Briefly, a total of 64 tap water samples (17 samples in summer, 24 samples in winter and another 23 samples in spring during 2015–2016) were collected from 13 drinking water treatment works in 8 counties (or county-level-cities) in Jinhua Region, which located in the middle of Zhejiang Province (S-Fig. 1) and has a typical sub-tropical monsoon climate. Each water works has its own reservoir(s) for source

Linearly and log-linearly regression models

Linear (model 1–3) and log-linear (model 4–6) models for THMs prediction were generated with the training samples (No.1–51) in S-Table 1. The results of the F test for the generated models are shown in Table 1. For Model 1, 3, 4 and 6, there are 2 water quality parameters, pw equals 2, and the corresponding n-pw-1 value is 48; while for Model 2 and 5, both of them only have one water quality parameter, so the pw and n- pw −1 value are 1 and 49, respectively. According to the F test (Table 1), F

Conclusions

The present study first used RBF ANN and GRA-RBF ANN to predict THMs levels in real tap water. Results show that RBF ANN consistently displayed higher regression coefficient and better prediction accuracy as compared to linear/log-linear regression models. It suggests the powerful ability of RBF to capture the complicated non-linear relationship between the water quality parameters and THMs formation. Meanwhile, RBF ANN was allowed to further improve the prediction accuracy by adjusting the

CRediT authorship contribution statement

Huachang Hong: Funding acquisition, Writing - original draft. Zhiying Zhang: Investigation, Methodology. Aidi Guo: Investigation, Data curation. Liguo Shen: Investigation, Data curation, Formal analysis. Hongjie Sun: Data curation, Formal analysis. Yan Liang: . Fuyong Wu: . Hongjun Lin: Project administration, Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This study was financially supported by Public Welfare Project of the Science and Technology Department of Zhejiang Province (LGF18H260005), Foundation of Science and Technology Bureau of Jinhua, Zhejiang Province, China (No. 2014-3-030) and National Natural Science Foundation of China (Nos. 51978628, 51578509, 22076171). The authors would like to thank Baoliang Dai, Chouye Wu and Mingbang Huang for their technical support.

References (61)

  • J. Hur et al.

    Characterization of chromophoric dissolved organic matter and trihalomethane formation potential in a recently constructed reservoir and the surrounding areas - Impoundment effects

    J. Hydrol.

    (2014)
  • S.A. Iliyas et al.

    RBF neural network inferential sensor for process emission monitoring

    Control Eng. Pract.

    (2013)
  • L. Jin et al.

    QSPR study on normal boiling point of acyclic oxygen containing organic compounds by radial basis function artificial neural network

    Chemometr. Intell. Lab.

    (2016)
  • P. Kulkarni et al.

    Disinfection by-product formation following chlorination of drinking water: Artificial neural network models and changes in speciation with treatment

    Sci. Total Environ.

    (2010)
  • C.-K. Lin

    Radial basis function neural network-based adaptive critic control of induction motors

    Appl. Soft Comput.

    (2011)
  • J. Lin et al.

    Regression models evaluating THMs, HAAs and HANs formation upon chloramination of source water collected from Yangtze River Delta Region, China

    Ecotox. Environ. Safe.

    (2018)
  • W. Liu et al.

    Formation of disinfection byproducts in typical Chinese drinking water

    J. Environ. Sci. (China)

    (2011)
  • H.R. Mian et al.

    Prioritization of unregulated disinfection by-products in drinking water distribution systems for human health risk mitigation: A critical review

    Water Res.

    (2018)
  • T.K. Nissinen et al.

    Disinfection by-products in Finnish drinking waters

    Chemosphere

    (2002)
  • N.M. Peleato et al.

    Neural networks for dimensionality reduction of fluorescence spectra and prediction of drinking water disinfection by-products

    Water Res.

    (2018)
  • S.D. Richardson et al.

    Occurrence, genotoxicity, and carcinogenicity of regulated and emerging disinfection by-products in drinking water: a review and roadmap for research

    Mutation Res.

    (2007)
  • R. Sadiq et al.

    Disinfection by-products (DBPs) in drinking water and predictive models for their occurrence: a review

    Sci. Total Environ.

    (2004)
  • A. Selberg et al.

    Content and composition of natural organic matter in water of Lake Pitkjarv and mire feeding Kuke River (Estonia)

    J. Hydrol.

    (2011)
  • K.P. Singh et al.

    Artificial intelligence based modeling for predicting the disinfection by-products in water

    Chemometr. Intell. Lab.

    (2012)
  • J. Sohn et al.

    Disinfectant decay and disinfection by-products formation model development: Chlorination and ozonation by-products

    Water Res.

    (2004)
  • H. Sun et al.

    Formation of disinfection by-products during chlorination of organic matter from phoenix tree leaves and Chlorella vulgaris

    Environ. Pollut.

    (2018)
  • J. Teng et al.

    Different fouling propensities of loosely and tightly bound extracellular polymeric substances (EPSs) and the related fouling mechanisms in a membrane bioreactor

    Chemosphere

    (2020)
  • G. Tian et al.

    Green decoration materials selection under interior environment characteristics: A grey-correlation based hybrid MCDM method

    Renew. Sust. Energ. Rev.

    (2018)
  • H. Wong et al.

    Change-point analysis of hydrological time series using grey relational method

    J. Hydrol.

    (2006)
  • B. Wu et al.

    Hydrophobic organic compounds in drinking water reservoirs: Toxic effects of chlorination and protective effects of dietary antioxidants against disinfection by-products

    Water Res.

    (2019)
  • Cited by (78)

    View all citing articles on Scopus
    View full text