Partial least squares regression models to predict contaminant concentrations during high or low flow of coal mine‐affected rivers

Exploratory analysis of existing multivariate river water datasets can provide useful insights during river basin research and can be used to identify important environmental variables and data suitable for concentration prediction models. In this work, a large dataset pertaining to coal mining areas of the Fitzroy River Basin, Australia, was used to demonstrate principal component analysis and partial least squares regression (PLS) modelling. In this example, a strong association between variables confirmed that that sodium was a major ion responsible for electrical conductivity across this vast river basin (PC-1). Suspected effects of dilution, evapoconcentration, and the influence of anthropogenic inputs on concentrations of nitrogen, sulfate and dissolved metals were also elucidated (PC-2 and PC-3). PLS models of a Comet, Nogoa, Mackenzie Rivers-subset indicated turbidity, dissolved Fe, total Ni, Co and Mn concentrations were not as variable during high flow as during low flow. Conductivity, sulfate and sodium concentrations were negatively correlated (>|0.7|) with total suspended solids (TSS) and several total and dissolved metals during both river conditions. Dissolved Al and Fe had a strong inverse relationship with total Fe and total Co concentrations during high flow. These relationships can be investigated further during future targeted monitoring and analysis. This work provided detailed methodology for development of concentration predictions models for parameters of environmental interest. Specifically, TestSet validated PLS models for dissolved Al (± 5.6 μg/L) and TSS (± 4.6 mg/L), and random Cross Validation models for TSS concentration (± 4.0 mg/L) during low flow and (± 3.5 mg/L) during high flow were produced.