A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Uncertainty Assessment of Surface Water Salinity Using Standalone, Ensemble, and Deep Machine Learning Methods: A Case Study of Lake Urmia
https://orcid.org/http://orcid.org/0009-0005-8105-2658
Due to the inherent uncertainty in surface water quality modeling, point predictions of water quality parameters such as total dissolved solids (TDS) and electrical conductivity (EC) as measures of water salinity concentrations are deemed insufficient. Specifically, the scarcity of statistical information regarding factors, such as untraceable illegal digging of wells near lakes for non-essential use of agricultural water, the impact of wastewater wells, and alteration of water surface flows by upstream dams, highlights the significance of incorporating evolved uncertainty in salinity amounts. As a result, the determination of prediction intervals (PIs) for salinity modeling becomes crucial. The main objective of this study is to compare PIs for EC and TDS predictions at five different stations in Lake Urmia using three machine learning models: deep neural network, i.e., deep multilayer perceptron (DMLP); standalone machine learning, i.e., support vector regression (SVR); and an ensemble learning model, i.e., random forest (RF), based on the Bootstrap method. A set of input variables, including pH, sodium adsorption ratio (SAR), total hardness (TH), and lake water level spanning over 16 years (195 months) from October 2005 to December 2021, was used for modeling. The results obtained from both the PIs and point prediction tasks demonstrated that the DMLP model with a customized cost function defined by the coverage width-based criterion (CWC) outperformed the SVR and RF models within the testing period. Additionally, the DMLP-based Bootstrap method achieved 71 and 67% lower CWC in predicting EC and 88 and 91% lower CWC in predicting TDS, respectively, compared to the SVR and RF models. It is also noteworthy that the inclusion of water level as an input parameter resulted in increased model accuracy. The study further revealed that the DMLP-based Bootstrap method performed better in predicting TDS compared to EC, with 76, 40, and 10% lower CWC in DMLP, SVR, and RF models, respectively.
Uncertainty Assessment of Surface Water Salinity Using Standalone, Ensemble, and Deep Machine Learning Methods: A Case Study of Lake Urmia
https://orcid.org/http://orcid.org/0009-0005-8105-2658
Due to the inherent uncertainty in surface water quality modeling, point predictions of water quality parameters such as total dissolved solids (TDS) and electrical conductivity (EC) as measures of water salinity concentrations are deemed insufficient. Specifically, the scarcity of statistical information regarding factors, such as untraceable illegal digging of wells near lakes for non-essential use of agricultural water, the impact of wastewater wells, and alteration of water surface flows by upstream dams, highlights the significance of incorporating evolved uncertainty in salinity amounts. As a result, the determination of prediction intervals (PIs) for salinity modeling becomes crucial. The main objective of this study is to compare PIs for EC and TDS predictions at five different stations in Lake Urmia using three machine learning models: deep neural network, i.e., deep multilayer perceptron (DMLP); standalone machine learning, i.e., support vector regression (SVR); and an ensemble learning model, i.e., random forest (RF), based on the Bootstrap method. A set of input variables, including pH, sodium adsorption ratio (SAR), total hardness (TH), and lake water level spanning over 16 years (195 months) from October 2005 to December 2021, was used for modeling. The results obtained from both the PIs and point prediction tasks demonstrated that the DMLP model with a customized cost function defined by the coverage width-based criterion (CWC) outperformed the SVR and RF models within the testing period. Additionally, the DMLP-based Bootstrap method achieved 71 and 67% lower CWC in predicting EC and 88 and 91% lower CWC in predicting TDS, respectively, compared to the SVR and RF models. It is also noteworthy that the inclusion of water level as an input parameter resulted in increased model accuracy. The study further revealed that the DMLP-based Bootstrap method performed better in predicting TDS compared to EC, with 76, 40, and 10% lower CWC in DMLP, SVR, and RF models, respectively.
Uncertainty Assessment of Surface Water Salinity Using Standalone, Ensemble, and Deep Machine Learning Methods: A Case Study of Lake Urmia
https://orcid.org/http://orcid.org/0009-0005-8105-2658
Due to the inherent uncertainty in surface water quality modeling, point predictions of water quality parameters such as total dissolved solids (TDS) and electrical conductivity (EC) as measures of water salinity concentrations are deemed insufficient. Specifically, the scarcity of statistical information regarding factors, such as untraceable illegal digging of wells near lakes for non-essential use of agricultural water, the impact of wastewater wells, and alteration of water surface flows by upstream dams, highlights the significance of incorporating evolved uncertainty in salinity amounts. As a result, the determination of prediction intervals (PIs) for salinity modeling becomes crucial. The main objective of this study is to compare PIs for EC and TDS predictions at five different stations in Lake Urmia using three machine learning models: deep neural network, i.e., deep multilayer perceptron (DMLP); standalone machine learning, i.e., support vector regression (SVR); and an ensemble learning model, i.e., random forest (RF), based on the Bootstrap method. A set of input variables, including pH, sodium adsorption ratio (SAR), total hardness (TH), and lake water level spanning over 16 years (195 months) from October 2005 to December 2021, was used for modeling. The results obtained from both the PIs and point prediction tasks demonstrated that the DMLP model with a customized cost function defined by the coverage width-based criterion (CWC) outperformed the SVR and RF models within the testing period. Additionally, the DMLP-based Bootstrap method achieved 71 and 67% lower CWC in predicting EC and 88 and 91% lower CWC in predicting TDS, respectively, compared to the SVR and RF models. It is also noteworthy that the inclusion of water level as an input parameter resulted in increased model accuracy. The study further revealed that the DMLP-based Bootstrap method performed better in predicting TDS compared to EC, with 76, 40, and 10% lower CWC in DMLP, SVR, and RF models, respectively.
Uncertainty Assessment of Surface Water Salinity Using Standalone, Ensemble, and Deep Machine Learning Methods: A Case Study of Lake Urmia
Iran J Sci Technol Trans Civ Eng
Raheli, Bahareh (author) / Talebbeydokhti, Nasser (author) / Saadat, Solmaz (author) / Nourani, Vahid (author)
2024-04-01
19 pages
Article (Journal)
Electronic Resource
English
| British Library Conference Proceedings | 2013