Predicting the Oxidant Demand in Full-Scale Water Treatment using an Artificial Neural Network: Uncertainty and Sensitivity Analysis

Lluís Godo-Pla, Pere Emiliano, Fernando Valero, Manel Poch, Gürkan Sin, Hèctor Monclús*

*Corresponding author for this work

Research output: Contribution to journalJournal articleResearchpeer-review

Abstract

Drinking Water Treatment Plants face changes in raw water quality and quantity and the treatment needs to be adjusted accordingly to produce the best water quality at the minimum environmental cost. The amount of data generated along drinking water treatment plants allows developing data-based models like artificial neural networks that are able to predict operational parameters and can be incorporated into environmental decision support systems. In the present study, an artificial neural network is developed for predicting the potassium permanganate demand at the inlet of a full-scale Drinking Water Treatment Plant. A systematic methodology is carried out for outlier detection and removal from the original dataset. Afterwards, model parameters estimation, uncertainty and sensitivity analysis is reported to assess prediction quality and uncertainty of the models. Bootstrap method was used for parameter estimation, and uncertainty of the inputs onto the model outputs was propagated using a Monte Carlo scheme. Several sensitivity analysis methods were evaluated to understand the contribution of the inputs on the output of the models, and this was in accordance with the knowledge of the process and other studies found in the literature. The selected architecture consisted of a feed-forward multi-layer perceptron with four inputs and one node in the hidden layer with a sigmoid activation function. The mean absolute error of the resulting model is 0.128 mg·L-1, which was considered acceptable by the DWTP operators. The resulting model provided good results in terms of replicative, predictive and structural performance and is to be used for supporting decision-making in the daily operation of the plant.
Original languageEnglish
JournalProcess Safety and Environmental Protection
Volume125
Pages (from-to)317-327
ISSN0957-5820
DOIs
Publication statusPublished - 2019

Keywords

  • Artificial neural network
  • Uncertainty
  • Sensitivity
  • Water treatment
  • Oxidation
  • Permanganate

Cite this

Godo-Pla, Lluís ; Emiliano, Pere ; Valero, Fernando ; Poch, Manel ; Sin, Gürkan ; Monclús, Hèctor. / Predicting the Oxidant Demand in Full-Scale Water Treatment using an Artificial Neural Network: Uncertainty and Sensitivity Analysis. In: Process Safety and Environmental Protection. 2019 ; Vol. 125. pp. 317-327.
@article{3147c542649b406ba91289c6775b02d9,
title = "Predicting the Oxidant Demand in Full-Scale Water Treatment using an Artificial Neural Network: Uncertainty and Sensitivity Analysis",
abstract = "Drinking Water Treatment Plants face changes in raw water quality and quantity and the treatment needs to be adjusted accordingly to produce the best water quality at the minimum environmental cost. The amount of data generated along drinking water treatment plants allows developing data-based models like artificial neural networks that are able to predict operational parameters and can be incorporated into environmental decision support systems. In the present study, an artificial neural network is developed for predicting the potassium permanganate demand at the inlet of a full-scale Drinking Water Treatment Plant. A systematic methodology is carried out for outlier detection and removal from the original dataset. Afterwards, model parameters estimation, uncertainty and sensitivity analysis is reported to assess prediction quality and uncertainty of the models. Bootstrap method was used for parameter estimation, and uncertainty of the inputs onto the model outputs was propagated using a Monte Carlo scheme. Several sensitivity analysis methods were evaluated to understand the contribution of the inputs on the output of the models, and this was in accordance with the knowledge of the process and other studies found in the literature. The selected architecture consisted of a feed-forward multi-layer perceptron with four inputs and one node in the hidden layer with a sigmoid activation function. The mean absolute error of the resulting model is 0.128 mg·L-1, which was considered acceptable by the DWTP operators. The resulting model provided good results in terms of replicative, predictive and structural performance and is to be used for supporting decision-making in the daily operation of the plant.",
keywords = "Artificial neural network, Uncertainty, Sensitivity, Water treatment, Oxidation, Permanganate",
author = "Llu{\'i}s Godo-Pla and Pere Emiliano and Fernando Valero and Manel Poch and G{\"u}rkan Sin and H{\`e}ctor Moncl{\'u}s",
year = "2019",
doi = "10.1016/j.psep.2019.03.017",
language = "English",
volume = "125",
pages = "317--327",
journal = "Process Safety and Environmental Protection",
issn = "0957-5820",
publisher = "Elsevier",

}

Predicting the Oxidant Demand in Full-Scale Water Treatment using an Artificial Neural Network: Uncertainty and Sensitivity Analysis. / Godo-Pla, Lluís; Emiliano, Pere; Valero, Fernando; Poch, Manel; Sin, Gürkan; Monclús, Hèctor.

In: Process Safety and Environmental Protection, Vol. 125, 2019, p. 317-327.

Research output: Contribution to journalJournal articleResearchpeer-review

TY - JOUR

T1 - Predicting the Oxidant Demand in Full-Scale Water Treatment using an Artificial Neural Network: Uncertainty and Sensitivity Analysis

AU - Godo-Pla, Lluís

AU - Emiliano, Pere

AU - Valero, Fernando

AU - Poch, Manel

AU - Sin, Gürkan

AU - Monclús, Hèctor

PY - 2019

Y1 - 2019

N2 - Drinking Water Treatment Plants face changes in raw water quality and quantity and the treatment needs to be adjusted accordingly to produce the best water quality at the minimum environmental cost. The amount of data generated along drinking water treatment plants allows developing data-based models like artificial neural networks that are able to predict operational parameters and can be incorporated into environmental decision support systems. In the present study, an artificial neural network is developed for predicting the potassium permanganate demand at the inlet of a full-scale Drinking Water Treatment Plant. A systematic methodology is carried out for outlier detection and removal from the original dataset. Afterwards, model parameters estimation, uncertainty and sensitivity analysis is reported to assess prediction quality and uncertainty of the models. Bootstrap method was used for parameter estimation, and uncertainty of the inputs onto the model outputs was propagated using a Monte Carlo scheme. Several sensitivity analysis methods were evaluated to understand the contribution of the inputs on the output of the models, and this was in accordance with the knowledge of the process and other studies found in the literature. The selected architecture consisted of a feed-forward multi-layer perceptron with four inputs and one node in the hidden layer with a sigmoid activation function. The mean absolute error of the resulting model is 0.128 mg·L-1, which was considered acceptable by the DWTP operators. The resulting model provided good results in terms of replicative, predictive and structural performance and is to be used for supporting decision-making in the daily operation of the plant.

AB - Drinking Water Treatment Plants face changes in raw water quality and quantity and the treatment needs to be adjusted accordingly to produce the best water quality at the minimum environmental cost. The amount of data generated along drinking water treatment plants allows developing data-based models like artificial neural networks that are able to predict operational parameters and can be incorporated into environmental decision support systems. In the present study, an artificial neural network is developed for predicting the potassium permanganate demand at the inlet of a full-scale Drinking Water Treatment Plant. A systematic methodology is carried out for outlier detection and removal from the original dataset. Afterwards, model parameters estimation, uncertainty and sensitivity analysis is reported to assess prediction quality and uncertainty of the models. Bootstrap method was used for parameter estimation, and uncertainty of the inputs onto the model outputs was propagated using a Monte Carlo scheme. Several sensitivity analysis methods were evaluated to understand the contribution of the inputs on the output of the models, and this was in accordance with the knowledge of the process and other studies found in the literature. The selected architecture consisted of a feed-forward multi-layer perceptron with four inputs and one node in the hidden layer with a sigmoid activation function. The mean absolute error of the resulting model is 0.128 mg·L-1, which was considered acceptable by the DWTP operators. The resulting model provided good results in terms of replicative, predictive and structural performance and is to be used for supporting decision-making in the daily operation of the plant.

KW - Artificial neural network

KW - Uncertainty

KW - Sensitivity

KW - Water treatment

KW - Oxidation

KW - Permanganate

U2 - 10.1016/j.psep.2019.03.017

DO - 10.1016/j.psep.2019.03.017

M3 - Journal article

VL - 125

SP - 317

EP - 327

JO - Process Safety and Environmental Protection

JF - Process Safety and Environmental Protection

SN - 0957-5820

ER -