Author(s): Daan Bertels; Patrick Willems

Linked Author(s): Daan Bertels, Patrick Willems

Keywords: Salinity; Long-term simulation; Physics-informed machine learning; Water quality

Abstract: Accurate, fast and robust models of water systems are crucial to develop optimal management strategies that assure the availability of good quality water. This is especially true in light of changing boundary conditions such climate change, land use changes, changing demand or operations by competing water users. Optimization of management strategies or statistical analysis of impacts of adaptation measures using long-term simulations require computationally very efficient models. Physics-based models are typically not sufficiently efficient for such tasks. These detailed models also typically require a lot of information about the study area, which is not always available, and they rely on assumptions about the system’s behaviour which are not always fully valid. Machine learning (ML) techniques are often explored as computationally very efficient alternatives or surrogates for detailed physics-based models. In addition to their computational efficiency, ML techniques have the ability to discover additional dependencies between in- and outputs that are not known to the model developer or used in the physics-based models. Some drawbacks are that ML models can produce results that are physically inconsistent and that they may make unreliable predictions under boundary conditions not present in the training dataset. The presented research proposes a novel ML-model architecture that is designed for predicting time varying pollutant concentrations in a surface water system. The proposed concept uses memory cells to track the system’s state, based on the Long Short-Term Memory (LSTM) architecture. Those memory values can be interpreted as water volumes and pollutant masses in the system. By enforcing conservation of both volume and mass trough the design of the model’s architecture, the problem of physical inconsistency may be reduced and the trained models may be better suited for out-of-sample scenarios. This integration of scientific knowledge in ML is a practice that is rapidly picking up momentum in the research community. The novel methodology is demonstrated for a test case in Belgium to model salinity in the Albert Canal, a shipping canal that is also a major surface drinking water source. A finite volume method (FVM) model is used to generate training and testing data. The new physics-informed ML model is able to accurately predict the start and duration of increasing concentrations at the inlet point, which is relevant for the water production, at only a fraction of the calculation time of the detailed FVM model. The developed model can be used for long-term simulations to statistically analyse the production plant’s reliability under climate change and changing management of the shipping canal, for example installation of additional pumping stations and new lock complexes.

DOI: https://doi.org/10.3850/IAHR-39WC252171192022697

Year: 2022