Prof. dr. ir. Ingmar Nopens - Ghent University
Prof. dr. ir. Elena Torfs – Université Laval
- For the optimal operation of water production, digital twins or the real-time use of models would represent a powerful tool to monitor water quality and improve efficiency. Their potential to optimise chemicals and energy consumption is nevertheless still unexploited in the industry.
- Many process units in water production are still not operated efficiently, and the information provided by the manufacturers does not always allow to optimise their performance. For instance, ion exchange resins are affected by fouling and their suboptimal regeneration and associated waste disposal constitutes one of the major operating costs in terms of utilised chemicals.
- A powerful model capable of predicting water quality and resin saturation would allow to reduce the energetic and material impact of these processes, maximising the environmental and economic sustainability of water production.
- Development and validation of a dynamic model with sufficient predictive power to be used as digital twin for the operation of water production processes, such as ion exchange (IX) treatment trains.
- Test of a hybrid approach, combining a mechanistic and a data-driven model, to improve the predictive power of the fully mechanistic IX model.
- Application of a surrogate model in a model-based control algorithm, such as model predictive control (MPC) aiming for the optimisation of ion exchange regeneration frequency and the reduction of consumed chemicals.
- Mechanistic models (based on differential equations) provide a mathematical representation of the system through a predefined model structure. The model parameters will be determined from literature or by estimation from data covering different experimental conditions (model calibration).
- A sensitivity analysis will determine which parameters need be calibrated for both ion exchange and regeneration processes, based on the model structure.
- Unaccounted, complex phenomena such as resin fouling limit the predictive power of mechanistic models. Furthermore, dynamic models often involve the solution of partial differential equations (PDEs), which is computationally intensive and renders its use for control purposes difficult. For these reasons, a data-driven modelling approach will be proposed.
- Data-driven models help explain relationships in available process data and avoid the increase in complexity of mechanistic models, but the limited parameter interpretability and the inability of the model to extrapolate to unseen conditions restricts a full data-driven approach. A hybrid model, combining both paradigms, represents a promising framework.
- A hybrid approach is a trade-off between data and expert knowledge with the benefits and shortcomings of both types. Artificial neural networks (ANNs) will be trained in series or parallel to model the existing knowledge gaps.
- To enable its use in control algorithms, a surrogate model will be proposed where a data-driven model will fit the mechanistic counterpart with the aim of reducing the real-time computational requirements. Such a powerful model can be used as a soft sensor to control the ion exchange regeneration.