Machine learning-based optimal design of groundwater pollution monitoring network.

It is an important task of environmental management to design groundwater pollution monitoring network (GPMN) to find out the occurrence of pollution events and carry out remediation in time. However, there are many uncertain factors in the process of designing GPMN, which affect the GPMN design result. In the process of applying the Monte Carlo method for uncertainty analysis, groundwater numerical simulation model may be utilized thousands of times, which results in a huge computational load. In order to overcome this disadvantage, a machine learning (ML)-based surrogate model is constructed with Kriging method, to replace the computational simulation model under uncertainty of pollution sources and parameters. The 0-1 integer programming optimization model is constructed to maximally cover serious polluted area to detect the occurrence of groundwater pollution in time. The optimal design framework of GPMN based on proposed ML algorithm was applied in a domestic landfill in Baicheng City, China. The results showed that the ML-based surrogate model has a great fitness with the groundwater solute transport simulation model. The optimal results of GPMN indicated that monitoring wells should be mainly placed at the downstream of the leachate equalization basin. If more wells are allowed to be placed, part of wells could be placed at the downstream of the landfill. Moreover, the area where the pollution plumes of landfill site meet that of leachate equalization basin should be set as the key monitoring objective. Verification and comparison showed that the pollutant detection rate of the optimal layout scheme is far higher than random layout schemes, which proves the reliability of the ML-based optimal design scheme of GPMN.