|M.Sc Student||Opher Tamar|
|Subject||Development of a Data Driven Model for Estimation of|
Pollutant Levels in Highway Stormwater Runoff
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Eran Friedler|
|Full Thesis text|
Pollutants accumulated on road pavement during dry periods are washed off the road surface with runoff water during rainfall events, presenting a potentially hazardous non-point source of pollution. Estimation of pollutant loads in these runoff waters is required for developing mitigation and management strategies, yet the numerous factors involved and their complex interconnected influences make straightforward assessment almost impossible. DDM (data driven modeling) techniques have lately been used in environmental research and have shown very good results.
In this research MT (model tree) and GA (genetic algorithm) methodologies were combined and a coupled model was developed for EMC (event mean concentration) estimation of selected highway runoff pollutants. EMCs of CrTotal, PbTotal, ZnTotal, TOC and TSS were modeled, using different combinations of explanatory variables. Of the large collection of known influencing factors, five of the most significant ones were chosen to function as potential explanatory variables for the models: annual average daily traffic (AADT), antecedent dry period (ADP), rainfall volume, maximum rainfall intensity and antecedent rainfall volume. Contrary to traditional MTs explanatory variables in this implementation were represented by non-linear equations.
Models were trained and verified using a comprehensive dataset of runoff events, monitored in various highway sites in California, USA. For each pollutant the set of variables resulting in the best performing model was selected. Estimation ability of the models in terms of correlation between estimated and actual values of both training and verification data was mostly higher than previously reported values. PbTotal was modeled with an outcome of R2 of 0.95 on training data and 0.43 on verification data. The developed model for TOC achieved R2 values of 0.91 and 0.49 on training and verification data respectively. Sensitivity analysis was carried out with satisfactory results.
AADT was found to be the most significant influencing factor, being the only attribute included in all five models. Rainfall volume was the second-most influencing parameter, included in four of the five models, followed by ADP, which was used as an explanatory variable by three of the models.
Generally, the models do a good job of providing the expected EMC of each of the five modeled pollutants, given input data related to highway traffic, dry period, current storm and previous storm. Further development is needed in fine-tuning the integrated mathematical expressions and the range limits of their coefficients, if these models are to be used for practical applications of highway and BMP planning.