Watershed-scale water environmental capacity estimation assisted by machine learning

Abstract

Water environmental capacity (WEC), the maximum amount of contaminants that a water body system can take without unacceptable impact to water quality in the system, is an important index for the managements of water resources and environmental quality. Here we proposed a machine learning-assisted approach that can be used to estimate watershed-scale WEC. In the approach, a process-based model was used to simulate contaminant concentrations at monitoring or critical river locations in response to contaminant inputs in the watershed, while an artificial neural network (ANN) as a machine learning method was trained to link the contaminant inputs in the watershed with the contaminant concentrations at the critical locations. From the linkages, a watershed-scale WEC that meets water quality constraints was obtained using a global optimization method. The integration of ANN in the WEC estimation is computationally efficient that can avoid exhaustive search of WEC using the process-based model only, especially in a complex river network system. Maozhou River watershed located at Shenzhen City, Southeast China, was used as an example to illustrate the approach with ammonium as an example contaminant. The obtained optimal WEC value varied with different water quality constraints and input distributions. The approach can be used to estimate WEC in the watershed with and without pre-existence contaminant inputs by optimizing the design of new inputs and their distribution. The results had an important implication for future watershed-scale water environmental management.