This blog post highlights some of the key features and findings of a recently published paper from Water Science & Technology.

Hybrid modelling of nitrogen removal by biofiltration using high-frequent operational data

https://doi.org/10.2166/wst.2024.293

Marcello Serrao, Vincent Jauzein, Ilan Juran, Bruno Tassin, Peter Vanrolleghem

How can process modelling optimize biofiltration systems used for wastewater treatment?

Optimization of wastewater treatment processes is a highly dynamic topic for the water industry – where operators are forced to manage a growing urbanization and on-going climate change at the same time as reducing the economic and ecological impact of water treatment, whilst still guaranteeing compliance to stringent water quality norms for the discharge of treated water. Innovative processing systems have been designed as a result, such a biofiltration, that are characterized by adaptability and resilience to changing conditions. The academic world as well, is actively contributing to process optimization. Current research topics involve the application of advanced data analysis tools, process modelling and artificial intelligence for systems design, process control and risks assessments. The trend in the wastewater sector tends to navigate towards digital twins of systems (or even entire treatment plants), for which reliable and precise process models are a requisite.

In the case of wastewater treatment by biofiltration, in which a dense ‘fixed-culture’ biomass of micro-organisms is exploited in a thin biofilm layer to transform pollutants into benign byproducts, the process models need to describe these biofilm-specific processes (e.g., diffusion, competition for space among several types of biomasses and regular backwashing to prevent filter clogging). Such biofilm models contain many parameters, are considered complex to fit the ever-changing reality (frequent recalibration) and suffer from inherent numerical challenges that require more computing power. All this makes them less suitable for real-time applications – such as in digital twins - where model updates are needed in real time with a short time horizon of action. Conversely, data-driven models that use artificial intelligence techniques – such as machine learning algorithms - to find patterns in data are computationally very fast, making them very interesting for real-time applications.

An active research topic in the water sector – to which this study adheres to - focuses therefore on the development of integrated or hybrid models allowing to harness the powers of two types of process models; one based on deterministic and/or mechanistic models simulating process behavior; and a second type of model based on pure data analyses and machine learning (AI/ML) approaches. The expected overall output is an improvement of the speed and accuracy of predicted key water quality variables.

The scientific research that let to this paper was groundbreaking since for the first time a hybrid model was developed for an operational biofiltration system in use in a municipal wastewater treatment facility. It is also the first hybrid model that was calibrated and trained on high-frequent operational data collected by inline sensors and supported by daily laboratory sampling analyses over a long period.  Commonly, most modelling studies often use experimental data from a pilot-study or otherwise controlled environment.

The hybrid model developed in this study combines a mechanistic biofilm reactor model in parallel with a machine learning model capable of running in near real-time in a python environment controlling both types of models interactively: a dynamic biofilm model of a submerged biofilter reactor (that includes proven equations for understood process behavior of micro-organisms such as growth and decay rates) and a data-driven neural network model that corrects for any missed (or misinterpreted) process dynamics.

The results obtained with simulations over the training and testing periods establish the validity of the hybrid model. The trained data-driven model (a feed-forward neural network) can pick-up ‘hidden’ dynamic information that corrects the mechanistic model's residual error. This shows that mechanistic models, particularly those developed for biofiltration, require assumptions that fail to include all process dynamics, some of which can be ‘recovered’ by the data-driven component of the hybrid model. The validation error of the hybrid model is three times smaller than the error of the mechanistic-only model. This is confirmed by the strong performance indicators, which underline the capability of the model to achieve under unseen conditions.

You may ask yourself, what are the future perspectives? Well, altogether, this study indicates that hybrid models make more reliable predictions of key water quality variables, thus improving confidence in the model's output. This then supports future operational applications of hybrid biofilm models, such as digital twins.

However, considering the highly dynamic conditions observed in wastewater treatment plants, the considerable measurement errors often observed in the monitoring data, and the sensitivity of the model calculations to outliers and faults in the input data, the model performance could further benefit from more intense (and automated) data preparation and reconciliation processes, as well as more data from longer training and testing periods. This would then pave the way to a future development of a digital twin for process control in biofiltration systems. One such application that is being studied now by the authors is how we can take advantage of the developed hybrid model and place it in a digital twin central to a new controller for dosing of chemical reagents optimizing nitrogen removal in a biofilm reactor.

This research was supported by SIAAP (Paris, France) as part of the Mocopée Research Program (innEAUvation.fr) that includes partnerships with modelEAU - Université Laval (Québec, Canada), LEESU – ENPC (Champs sur Marne, France) and W-SMART (Paris, France).