Designing soil contamination monitoring network in petroleum refineries by XGBoost weighting and geostatistical...

Petroleum refineries are deemed strategic industrial sectors that can release toxic materials to the environment and cause potential hazards. In this regard, designing and installation of soil contamination monitoring networks at petroleum refineries is a necessity. In this research, we designed an optimal monitoring network with maximum coverage and minimum number of monitoring boreholes. The main regarded parameters are the groundwater contamination history, the location of effective structures, the location of flare stacks and the soil texture. In addition, the soil contamination was calculated based on previous contamination of the soil at the sampling points by the Entropy Weighting Model. It was employed with other parameters to estimate the soil contamination across the site. The Machine Learning method of XGBoost was implemented for estimating and assigning priority for every point of the site. To achieve the optimal network in the optimization program, four parameters were regarded including (a) the optimal value of the optimization program's objective function, (b) the number of Advance Zero-half cuts of the Cut Generation algorithm, (c) the consumed time, and (d) the optimal boreholes number of the network corresponding with different effective contamination detection radius. The network was designed by generalized Maximal Covering Location Problem and for optimizing it, the advantages of Mixed-Integer Linear Programming method were used. To evaluate the applicability of the method, it has been developed and implemented in a refinery in the south of Iran. 92.84% of XGBoost estimation accuracy, the optimal number of 113 and the effective contamination detection radius of 160 m were obtained for boreholes of the network. To investigate the efficiency of the model, a new Regret function has been defined. Furthermore, sensitivity analysis of the parameters and feature importance analysis of XGBoost both showed that the main parameter of the model was the location of effective structures.