The water sector is on the search for ways to optimize the treatment of water and by-products to allow reducing the energy consumption and usage of chemical reagents while continuing to comply with the regulatory requirements for the produced water and biomass quality. Important drivers are the evolving human and environmental health regulations that will include emerging micro-pollutants and pathogens, the reduction of the dependency on fossil fuels as an energy source, and the concerns about environmental impact of operations (e.g. energy usage for pumps). Added to this are global pressures caused by an ongoing urbanization and the effects of climate change.
The development of innovative treatment systems for the water industry has created systems with a higher reactivity and operational flexibility requiring a smaller spatial footprint, which is an important feature for utilities located in densely urbanized settings. At the same time, the degrees of process control have increased due to the operational complexity. Water processing consists of several successive physical, biological and chemical processes that exhibit non-linear and non-stationary behaviour. This requires advanced automated process control to assure overall performance and compliance. In addition, AI and its machine learning (ML) methodologies create large expectancies partly thanks to the increasing availability of “big data” in the water sector.
Process Modelling
In the last few decades, experiences have been gained and lessons learned with the application of smart systems in the water and sanitation sector. This is observed in particular in the fields of in-line monitoring of flowrates and of qualitative process performance indicators used by automatic controllers. In addition, the use of computer-aided mathematical modelling to simulate and gain understanding of physical, chemical and microbial processes is advancing at a fast pace. Numerical mathematical models based on empirical and phenomenological principles are providing a deeper insight in the biochemical and hydraulic processes. Since the late 90’s, mechanistic (‘knowledge-based’) models have found their way from research institutes to utilities in order to support performance evaluation and decision-making. More and more now, water treatment process models are recognized as important decision support tools in the water and sanitation sector for process design, scenario analysis, impact assessment, optimization of treatment and process control, as well as for educational purposes.
Process Knowledge-based Models
One of the most widely accepted knowledge-based models are the family of Activated Sludge Models (ASM) that simulate the removal of carbon, nitrogen and phosphorus compounds during water treatment. The ASM1 model (1987) was well-received in the scientific community and served as the basis for further model development. The matrix notation that was introduced with the ASM1 model combines the soluble and particulate components affected by each of the simulated biological processes with their reaction rate terms. This matrix proved to be very practical to communicate about these complex models. The introduction of the ASM model was indeed of great importance, providing researchers and practitioners with a standardised set of basis models. Many process models have since then been developed based on a modified ASM model to allow simulating new technologies and the growing interest in studying environmental impact, such as with the complete nitrogen cycle and greenhouse gas emissions. Generally, these models have been calibrated with short-term experimental data in pilot set-ups and have reached reasonable performance. The greater complexity of models for novel treatment technologies makes their parameterization, calibration, and validation much more laborious, requiring more dedicated measurement data. This can become a bottleneck to model application, especially in an operational real-time process control.
Hybrid Models
Alternatively, model-hybridization is a process that combines a mechanistic model, which integrates relevant knowledge about processes, with a Machine Learning model that increases the accuracy of estimates by including information about poorly described sub-processes. Hybridization of mechanistic models (MM) with data-driven models (DDM) allows taking advantage of the strong points of each model type: predictions of water quality variables are based on the fundamental understanding of the physical-chemical-biological processes that take place within a system. Data-driven Machine Learning models describe a system only based on the information extracted from the data. Such DDM models have strong interpolation strengths and are much faster in computation, which makes them very attractive for real-time process control applications. The data-driven model learns unknown relationships from the data which can help to correct for process dynamics missed by the mechanistic model. This will improve overall model simulation quality, especially for interpolation purposes. Research on the development of hybrid models describing activated sludge processes has improved model performance and supported process monitoring and control. This has reportedly created a more stable and reliable process control and an overall increase in the understanding of the hydraulic, chemical, and biological processes taking place during the transport and treatment of water.
Innovation on Hybrid Models for Biofilms
More recently, the hybrid modelling approach has been applied for the first time to biofilm systems for water treatment. The bio-physical-chemical processes taking place within a biofilm, which is a thin layer composed of micro-organisms responsible for the degradation of pollutants into benign materials, make for complex mathematical equations. These biofilter mechanistic models require relatively high computing power and modelling guidelines have not yet been well established within the scientific community. Fortunately, change is coming gradually, for example, due to the efforts of the InnEAUvation Program (SIAAP, France) that promotes applied research on wastewater technologies. Within that framework, the outcome of a recent PhD study at the LEESU (Ecole des ponts et chaussées - France) in co-direction with modelEAU (Université Laval - Canada) on the hybrid modelling of biofilm systems in wastewater treatment as presented at the Watermatex 2023 conference in Québec (Canada) shows that the output of the hybrid model is much closer to the observations for the produced water quality when compared to the results of the mechanistic model only. The developed hybrid model integrates a data-driven model to significantly reduce the prediction mean error of the biofilter model with 80% during training and 75% during testing. It is then applied as an integral part of a Hybrid Model Predictive Control system to achieve improved process control of this intensive process. Evaluation indicators confirm the validity of the hybrid model showing that the data-driven component captured sufficient residual information to compensate for the inaccuracy of the mechanistic model.
Data Science Training
An important subject of focus is the training (or re-schooling) of water engineers and operators in the fields of statistical analysis, data science and process modelling. For a successful digital transformation, the water sector needs to create ample opportunities for organizational and cultural changes that allow for cross-disciplinary training in areas of water management and data science. Participation in online platforms where modelling engineers and (water) data scientists can exchange knowledge and best-practices on the design of hybrid models is growing, for example with the recent establishment of the IWA Working Group on Hybrid Modelling.
Take-away Messsages
The overall good results obtained during research studies so far on the application of hybrid models in water and wastewater treatment sector supports future applications in an operational context. Hybrid models allow reducing the size of model prediction errors and increase the confidence in the model’s performance so that it can be safely applied as a decision support tool and – in the near future - even as a model predictive process controller. The training of operators and engineers alike in the field of Data Science would be very beneficial to the rapid uptake by the water industry of hybrid modelling. The results of scientifc studies indicate that hybrid models have a higher benefit/cost ratio for solving complex problems, which is a key factor for process systems engineering.