Prediction of bisphenol A contamination in Canadian municipal wastewater

https://doi.org/10.1016/j.jwpe.2022.103304Get rights and content

Highlights

  • A framework for Bisphenol A (BPA) modeling at wastewater plants is proposed

  • Data from 12 plants are used to develop data-driven models for BPA prediction

  • Influencing factors of BPA removal are studied using network theory

  • The results imply that BPA can hardly be removed through primary treatment

  • Important factors for BPA removal at wastewater treatment plants are identified

Abstract

Bisphenol A (BPA) is one of the most common contaminants of emerging concerns (CECs), which pose a threat to human health. Conventional wastewater treatment plants (WWTPs) are considered as the major pathway of BPA entering the aqueous environment. To control and mitigate BPA contamination in the aquatic environment, predicting BPAs fate at WWTPs is critical. In this study, three machine learning models, including shared layer multi-task neural network (MLT-NN), genetic programming (GP), and extra trees (ET) are used to predict the effluent BPA concentration at twelve municipal WWTPs across Canada. Additionally, the theory of networks is adopted to analyze the interdependencies among the influencing factors of BPA removal. It is found that the proposed models can provide reasonable BPA effluent concentration predictions. They have advantages in alleviating data sparsity and imbalance, improving model interpretability, and measuring predictor importance, which is valuable for the modeling of BPA and many other CECs. The network analysis results imply there are moderate interdependencies among various influencing factors of BPA removal. Factors that significantly affect BPA effluent concentration and are thus important for BPA removal are identified. The results also show that BPA is unlikely to be removed at primary treatment plants, while BPA removal could be achieved through secondary or tertiary treatment. This study presents an integrated framework for the modeling and analysis of BPA at WWTPs, which can provide direct and robust decision support for the management of BPA as well as other emerging contaminants in municipal wastewater.

Introduction

Contaminants of emerging concerns (CECs), such as pharmaceuticals and personal care products (PPCPs), endocrine-disrupting compounds (EDCs), flame retardants (FRs), pesticides, and artificial sweeteners (ASWs), and their metabolites, are considered as a growing threat to the aqueous environment and public health [19], [27]. Conventional wastewater treatment plants (WWTPs) are designed to remove more commonly seen pollutants, such as organic matter, phosphorus, and nitrogen, not CECs [7]. As a result, WWTPs become one of the main pathways of the inductive release of CECs into the environment [29]. Understanding and predicting CECs' fate at WWTPs is of great importance to the mitigation of CEC-related risks.

Bisphenol A (BPA) is one of the most common CECs due to its massive use around the world. For the past few decades, BPA has been widely used as a raw material for manufacturing polycarbonate plastics and epoxy resins, which are used to produce daily consumer products such as water bottles, thermal paper, dental sealants, and medical equipment [14], [28]. BPA has been found to have an adverse impact on human health, being responsible for an increase in incidences such as cancer and hormonal imbalance [18]. Although various treatment methods (e.g., adsorption on activated carbon, ultrafiltration, biodegradation, and ozonation) have been proven to be effective for BPA removal, they are not available at most conventional WWTPs due to limited budget [5], [38]. Therefore, predicting BPA concentration in WWTP effluents is important for estimating the amount of BPA discharged into the aquatic environment. However, previous research on BPA modeling at WWTPs is very limited. Most previous studies on BPA are based on sample-by-sample analysis [6], [8], [16], [36], [40]. Lee and Peart [20] analyzed 36 Canadian wastewater influent/effluent sample pairs and reported that BPA in the influent can be removed during the treatment process at a median reduction rate of 68 %. Guerra et al. [14] investigated how parameters affect BPA occurrence, removal, and fate. More recently, several attempts have been made to investigate BPA distribution and modeling at large scales. For example, Gewurtz et al. [12] used a multimedia approach to assess spatial and temporal trends of BPA in the Canadian environment. Tong et al. [34] proposed a hybrid approach for BPA prediction in a reservoir that harvests rainfall water and acts as drinking and recreational water resource. To our knowledge, there are no previous studies on the prediction of BPA fate during municipal wastewater treatment processes. The decay and removal of BPA in wastewater are complex processes and it is hard to simulate such processes using conventional wastewater simulation models, which are typically process-based models (PBM) with limited capacity to capture complex relationships and are usually influenced by uncertain variables, such as pH and salting-out effects [15], [22].

Data-driven models (DDMs) have attracted much attention recently and have been successfully used as alternatives for conventional process-based models (PBMs) in the field of wastewater modeling [10], [25], [42]. DDMs have advantages over PBMs in capturing highly complex and nonlinear relationships, but the lack of data may be a major obstacle to developing DDMs [24], [39]. In the past decade, a national wastewater monitoring program in Canada that monitors chemical substances has made it possible to obtain a certain amount of laboratory data for emerging contaminants prediction. However, using normal data-driven modeling techniques to predict BPA in wastewater is still a daunting challenge due to the data imbalance and sparsity caused by limited laboratory resources, as well as the poor interpretability of traditional DDMs. To address such concerns, three well-customized DDMs including multitask shared layer neural network (MLT-NN), genetic programming (GP), and extra trees (ET) are introduced in this study. MLT-NN that can leverage useful information from other related learning tasks is used to alleviate the data sparsity and imbalance problem. Closed-form functions generated by GP and variable importance measures (VIM) derived from ET are used to better interpret the DDM results and investigate the impacts of different wastewater features on BPA effluent concentrations. In addition, theory of networks, which is known for visualizing interdependencies among features and has been widely applied in many fields, is adopted to study the interdependencies among different WWTP features and provide an insight into the influencing factors of BPA effluent concentration [23], [31].

The overall objective of this study is to establish an integrated framework for the prediction and evaluation of BPA removal at Canadian municipal WWTPs. This entails the following three tasks: (1) integrate data from different WWTPs and build data-driven models for the prediction of effluent BPA; (2) analyze how wastewater features (e.g., temperature, influent flow rate) affect the BPA effluent concentration; (3) assess the interdependencies among wastewater treatment features and further analyze the influencing factors of BPA effluent concentration. This study is the first attempt to develop DDMs for municipal wastewater BPA prediction. It can provide direct decision support for the removal and management of BPA through municipal WWTPs. It also provides an example of how existing challenges (i.e., data sparsity and imbalance, model interpretability, and feature importance measurement) can be addressed to predict contaminants of emerging concerns at WWTPs; the developed methodology can be extended to predict and manage various emerging contaminants at WWTPs in future.

Section snippets

Methodology

An integrated framework as shown in Fig. 1 is proposed for predicting BPA effluent concentration at municipal WWTPs in this study. The framework consists of two major parts: (1) DDMs for interval predictions of BPA effluent concentration; and (2) networks for feature independence analysis. For the DDM part, the selected MTL-NN, GP, and ET are introduced to solve the data imbalance problem associated with CEC data, generate a closed-form function for model interpretability, and measure variable

Study area and data collection

In this study, wastewater data at twelve anonymous WWTPs across Canada were collected from the database provided by Canada’s Chemical Management Plan. The WWTPs can be classified into different categories (i.e., primary, secondary, and tertiary) based on the treatment processes. Different WWTPs may have different treatment units and thus lead to different treatment efficiency. The characteristics of the twelve WWTPs are shown in Table 1. Sewage samples of raw influent and final effluent were

Deterministic prediction and model comparison

In this study, three DDMs (i.e., MLT-NN, ET, and GP) are utilized to predict effluent BPA concentration across the twelve selected wastewater treatment plants. One random sample from each WWTP is reserved for model validation due to the sample sparsity, while the rest of the samples are used for training. The model performance is evaluated by mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R2). Table A-1 in the Appendix presents the

Conclusions

In this study, an integrated framework was proposed for the prediction of BPA effluent concentrations at Canadian municipal WWTPs. The framework consists of two major parts: (1) DDMs for effluent BPA prediction; and (2) a network for feature dependencies analysis. Specifically, MLT-NN, GP, and ET models were applied to address the data sparsity problem and generate effluent BPA predictions. BPA influent concentration, seasonal information, influent flow rate, influent temperature, effluent

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.

Acknowledgments

This study is supported by the MacDATA Institute at McMaster University, Canada. We would like to express appreciation to fellows at MacDATA institute for their advice.

References (43)

  • A. Jhones dos Santos et al.

    Removal of bisphenol A from acidic sulfate medium and urban wastewater using persulfate activated with electroregenerated Fe2+

    Chemosphere

    (2021)
  • K.O. K’oreje et al.

    Occurrence and treatment of contaminants of emerging concern in the African aquatic environment: literature review and a look ahead

    J. Environ. Manag.

    (2020)
  • M. Murugananthan et al.

    Mineralization of bisphenol A (BPA) by anodic oxidation with boron-doped diamond (BDD) electrode

    J. Hazard. Mater.

    (2008)
  • B.L. Narayanan et al.

    A complex network theoretic approach for interdependence investigation: an application to radionuclide behavior in the subsurface

    Comput. Geosci.

    (2021)
  • Senthilselvan Natarajan et al.

    Resolving data sparsity and cold start problem in collaborative filtering recommender system using linked open data

    Expert Syst. Appl.

    (2020)
  • K.B. Newhart et al.

    Data-driven performance analyses of wastewater treatment plants: a review

    Water Res.

    (2019)
  • A. Saltelli

    Making best use of model evaluations to compute sensitivity indices

    Comput. Phys. Commun.

    (2002)
  • I.M. Sobol

    Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates

    Math. Comput. Simul.

    (2001)
  • X. Tong et al.

    Advancing prediction of emerging contaminants in a tropical reservoir with general water quality indicators based on a hybrid process and data-driven approach

    J. Hazard. Mater.

    (2022)
  • D. Wang et al.

    A machine learning framework to improve effluent quality control in wastewater treatment plants

    Sci. Total Environ.

    (2021)
  • L. Xu et al.

    Photocatalytic activity and mechanism of bisphenol A removal over TiO2−x/rGO nanocomposite driven by visible light

    Chem. Eng. J.

    (2018)
  • Cited by (0)

    View full text