Emulating process-based water quality modelling in water source reservoirs using machine learning

Process-based models are very efficient in simulating hydrodynamics and water quality in surface water bodies. However, their complex characteristics in terms of implementation, data requirements, and simulation time limit their application in regular drinking water source management. This study demonstrates the potential of a ML model (Long Short-Term Memory (LSTM)) as viable alternatives to process-based hydrodynamic and water quality models in water source management. Using meteorological and hydrological measurements, a hydrodynamic and water quality model was first calibrated to predict time series, profiles, and contours of water variables namely Escherichia coli (E. coli), faecal coliforms, zinc, and lead concentrations in the Brusdalsvatnet lake, which is the drinking water source for the city of Alesund in Norway. The results obtained were combined with the input data to train a suite of LSTM models to emulate the results achieved with the process-based modelling. The results indicate that ML models can conveniently reproduce the spatio-temporal evolution of water quality in the lake that is achievable with the process-based model, particularly when specific locations within the lake are of interest. Compared to R-2, NS and MSE ranges of 0.72-0.87, 0.68-0.85, and 0.21-0.44 achieved with the process-based model in the prediction of temperature in the lake, 0.78-0.95, 0.75-0.89, and 0.011-0.028 were respectively achieved in testing the LSTM models. Similar performance levels were achieved with the LSTM model in the prediction of Escherichia coli (E. coli), faecal coliforms, Zinc, and Lead concentrations at different depths in the lake. While setting up and training the LSTM model to emulate the process-based model simulations was very time-consuming, a validated model as developed in this study can offer an opportunity for real-time simulation of water quality in drinking water sources when integrated with cloud data transmission from field sensors.

Automated summary

This study explores the potential of a machine learning model (LSTM) as an alternative to process-based hydrodynamic and water quality models in water source management. By using meteorological and hydrological measurements, the study demonstrates that LSTM models can accurately predict the spatio-temporal evolution of water quality in surface water bodies. The results indicate that the LSTM models perform well and show specific advantages over process-based models in predicting water variables at specific locations.