Hybrid modelling of water resource recovery facilities: status and opportunities

Towards good modelling practice for parallel hybrid models for wastewater treatment processes

This study explores various approaches to formulating a parallel hybrid model (HM) for Water and Resource Recovery Facilities (WRRFs) merging a mechanistic and a data-driven model. In the study, the HM is constructed by training a neural network (NN) on the residual of the mechanistic model for effluent nitrate. In an initial experiment using the Benchmark Simulation Model no. 1, a parallel HM effectively addressed limitations in the mechanistic model's representation of autotrophic bacteria growth and the data-driven model's incapability to extrapolate. Next, different versions of a parallel HM of a large pilot-scale WRRF are constructed, using different calibration/training datasets and different versions of the mechanistic model to investigate the balance between the calibration effort for the mechanistic model and the compensation by the NN component. The HM can improve predictions compared to the mechanistic model. Training the NN on an independent validation dataset produced better results than on the calibration dataset. Interestingly, the best performance is achieved for the HM based on a mechanistic model using default (uncalibrated) parameters. Both long short-term memory (LSTM) and convolutional neural network (CNN) are tested as data-driven components, with a CNN HM (root-mean-squared error (RMSE) = 1.58 mg NO3-N/L) outperforming an LSTM HM (RMSE = 4.17 mg NO3-N/L).

Numerical modelling of surface aeration and N2O emission in biological water resource recovery

Biokinetic modelling of N2O production and emission has been extensively studied in the past fifteen years. In contrast, the physical-chemical hydrodynamics of activated sludge reactor design and operation, and their impact on N2O emission, is less well understood. This study addresses knowledge gaps related to the systematic identification and calibration of computational fluid dynamic (CFD) simulation models. Additionally, factors influencing reliable prediction of aeration and N2O emission in surface aerated oxidation ditch-type reactor types are evaluated. The calibrated model accurately predicts liquid sensor measurements obtained in the Lynetten Water Resource Recovery Facility (WRRF), Denmark. Results highlight the equal importance of design and operational boundary conditions, alongside biokinetic parameters, in predicting N2O emission. Insights into the limitations of calibrating gas mass-transfer processes in two-phase CFD models of surface aeration systems are evaluated.

A Review of Computational Modeling in Wastewater Treatment Processes

Wastewater treatment companies are facing several challenges related to the optimization of energy efficiency, meeting more restricted water quality standards, and resource recovery potential. Over the past decades, computational models have gained recognition as effective tools for addressing some of these challenges, contributing to the economic and operational efficiencies of wastewater treatment plants (WWTPs). To predict the performance of WWTPs, numerous deterministic, stochastic, and time series-based models have been developed. Mechanistic models, incorporating physical and empirical knowledge, are dominant as predictive models. However, these models represent a simplification of reality, resulting in model structure uncertainty and a constant need for calibration. With the increasing amount of available data, data-driven models are becoming more attractive. The implementation of predictive models can revolutionize the way companies manage WWTPs by permitting the development of digital twins for process simulation in (near) real-time. In data-driven models, the structure is not explicitly specified but is instead determined by searching for relationships in the available data. Thus, the main objective of the present review is to discuss the implementation of machine learning models for the prediction of WWTP effluent characteristics and wastewater inflows as well as anomaly detection studies and energy consumption optimization in WWTPs. Furthermore, an overview considering the merging of both mechanistic and machine learning models resulting in hybrid models is presented as a promising approach. A critical assessment of the main gaps and future directions on the implementation of mathematical modeling in wastewater treatment processes is also presented, focusing on topics such as the explainability of data-driven models and the use of Transfer Learning processes.

A full-scale operational digital twin for a water resource recovery facility-A case study of Eindhoven Water Resource Recovery Facility

Digital transformation for the water sector has gained momentum in recent years, and many water resource recovery facilities modelers have already started transitioning from developing traditional models to digital twin (DT) applications. DTs simulate the operation of treatment plants in near real time and provide a powerful tool to the operators and process engineers for real-time scenario analysis and calamity mitigation, online process optimization, predictive maintenance, model-based control, and so forth. So far, only a few mature examples of full-scale DT implementations can be found in the literature, which only address some of the key requirements of a DT. This paper presents the development of a full-scale operational DT for the Eindhoven water resource recovery facility in The Netherlands, which includes a fully automated data-pipeline combined with a detailed mechanistic full-plant process model and a user interface co-created with the plant's operators. The automated data preprocessing pipeline provides continuous access to validated data, an influent generator provides dynamic predictions of influent composition data and allows forecasting 48 h into the future, and an advanced compartmental model of the aeration and anoxic bioreactors ensures high predictive power. The DT runs near real-time simulations every 2 h. Visualization and interaction with the DT is facilitated by the cloud-based TwinPlant technology, which was developed in close interaction with the plant's operators. A set of predefined handles are made available, allowing users to simulate hypothetical scenarios such as process and equipment failures and changes in controller settings. The combination of the advanced data pipeline and process model development used in the Eindhoven DT and the active involvement of the operators/process engineers/managers in the development process makes the twin a valuable asset for decision making with long-term reliability. PRACTITIONER POINTS: A full-scale digital twin (DT) has been developed for the Eindhoven WRRF. The Eindhoven DT includes an automated continuous data preprocessing and reconciliation pipeline. A full-plant mechanistic compartmental process model of the plant has been developed based on hydrodynamic studies. The interactive user interface of the Eindhoven DT allows operators to perform what-if scenarios on various operational settings and process inputs. Plant operators were actively involved in the DT development process to make a reliable and relevant tool with the expected added value.

Modeling greenhouse gas emissions from riverine systems: A review

Despite the recognized importance of flowing waters in global greenhouse gas (GHG) budgets, riverine GHG models remain oversimplified, consequently restraining the development of effective prediction for riverine GHG emissions feedbacks. Here we elucidate the state of the art of riverine GHG models by investigating 148 models from 122 papers published from 2010 to 2021. Our findings indicate that riverine GHG models have been mostly data-driven models (83%), while mechanistic and hybrid models were uncommonly applied (12% and 5%, respectively). Overall, riverine GHG models were mainly used to explain relationships between GHG emissions and biochemical factors, while the role of hydrological, geomorphic, land use and cover factors remains missing. The development of complex and advanced models has been limited by data scarcity issues; hence, efforts should focus on developing affordable automatic monitoring methods to improve data quality and quantity. For future research, we request for basin-scale studies explaining river and land-surface interactions for which hybrid models are recommended given their flexibility. Such a holistic understanding of GHG dynamics would facilitate scaling-up efforts, thereby reducing uncertainties in global GHG estimates. Lastly, we propose an application framework for model selection based on three main criteria, including model purpose, model scale and the spatiotemporal characteristics of GHG data, by which optimal models can be applied in various study conditions.

Developing a hybrid model for predicting the reaction kinetics between chlorine and micropollutants in water

Understanding the reactivities of chlorine towards micropollutants is crucial for assessing the fate of micropollutants in water chlorination. In this study, we integrated machine learning with kinetic modeling to predict the reaction kinetics between micropollutants and chlorine in deionized water and real surface water. We first established a framework to predict the apparent second-order rate constants for micropollutants with chlorine by combining Morgan molecular fingerprints with machine learning algorithms. The framework was tuned using Bayesian optimization and showed high prediction accuracy. It was validated through experiments and used to predict the unreported apparent second-order rate constants for 103 emerging micropollutants with chlorine. The framework also improved the understanding of the structure-dependence of micropollutants' reactivity with chlorine. We incorporated the predicted apparent second-order rate constants into the Kintecus software to establish a hybrid model to profile the time-dependent changes of micropollutant concentrations by chlorination. The hybrid model was validated by experiments conducted in real surface water in the presence of natural organic matter. The hybrid model could predict how much micropollutants were degraded by chlorination with varied chlorine contact times and/or initial chlorine dosages. This study advances fundamental understanding of the reaction kinetics between chlorine and emerging micropollutants, and also offers a valuable tool to assess the fate of micropollutants during chlorination of drinking water.

Deep learning in wastewater treatment: a critical review

Modeling wastewater processes supports tasks such as process prediction, soft sensing, data analysis and computer assisted design of wastewater systems. Wastewater treatment processes are large, complex processes, with multiple controlling mechanisms, a high degree of disturbance variability and non-linear (generally stable) behavior with multiple internal recycle loops. Semi-mechanistic biochemical models currently dominate research and application, with data-driven deep learning models emerging as an alternative and supplementary approach. But these modeling approaches have grown in separate communities of research and practice, and so there is limited appreciation of the strengths, weaknesses, contrasts and similarities between the methods. This review addresses that gap by providing a detailed guide to deep learning methods and their application to wastewater process modeling. The review is aimed at wastewater modeling experts who are familiar with established mechanistic modeling approach, and are curious about the opportunities and challenges afforded by deep learning methods. We conclude with a discussion and needs analysis on the value of different ways of modeling wastewater processes and open research problems.

Physics-informed neural network-based serial hybrid model capturing the hidden kinetics for sulfur-driven autotrophic denitrification process

Sulfur-driven autotrophic denitrification (SdAD) is a biological process that can remove nitrate from low carbon/nitrogen (C/N) ratio wastewater. Although this process has been intensively researched, the mechanism whereby its intermediates (i.e., elemental sulfur and nitrite ions) are generated and accumulated remains elusive. Existing mathematical models developed for SdAD cannot accurately predict the intermediates in SdAD because of the incomplete knowledge of process kinetic resulting from changes in the environmental conditions and electron competition during SdAD. To address this limitation, we proposed a novel serial hybrid model structure based on a physics-informed neural network (PINN) to capture the dynamics of the process kinetics and predict the substrate concentrations in SdAD. In this study, we evaluated the model through numerical experiments and applied it to real case studies involving batch and continuous-flow reactor scenarios. By leveraging the PINN approach, the hybrid model yielded accurate predictions at both the state (i.e. substrate concentration) and kinetic levels in the numerical experiments and performed better than both mechanistic and purely data-driven models in the case studies. Furthermore, we used the trained hybrid model to design control strategies for SdAD and a novel integrated process involving SdAD and anammox for energy-efficient nitrogen removal. Finally, we discuss the advantages and application scope of the PINN-based hybrid model.

Predicting microbial water quality in on-site water reuse systems with online sensors

Widespread implementation of on-site water reuse is hindered by the limited availability of monitoring approaches that ensure microbial quality during operation. In this study, we developed a methodology for monitoring microbial water quality in on-site water reuse systems using inexpensive and commercially available online sensors. An extensive dataset containing sensor and microbial water quality data for six of the most critical types of disruptions in membrane bioreactors with chlorination was collected. We then tested the ability of three typological machine learning algorithms - logistic regression, support-vector machine, and random forest - to predict the microbial water quality as "safe" or "unsafe" for reuse. The main criteria for model optimization was to ensure a low false positive rate (FPR) - the percentage of safe predictions when the actual condition is unsafe - which is essential to protect users health. This resulted in enforcing a fixed FPR ≤ 2%. Maximizing the true positive rate (TPR) - the percentage of safe predictions when the actual condition is safe - was given second priority. Our results show that logistic-regression-based models using only two out of the six sensors (free chlorine and oxidation-reduction potential) achieved the highest TPR. Including sensor slopes as engineered features allowed to reach similar TPRs using only one sensor instead of two. Analysis of the occurrence of false predictions showed that these were mostly early alarms, a characteristic that could be regarded as an asset in alarm management. In conclusion, the simplest algorithm in combination with only one or two sensors performed best at predicting the microbial water quality. This result provides useful insights for water quality modeling or for applications where small datasets are a common challenge and a general advantage might be gained by using simpler models that reduce the risk of overfitting, allow better interpretability, and require less computational power.

Sensor setpoints that ensure compliance with microbial water quality targets for membrane bioreactor and chlorination treatment in on-site water reuse systems

Widespread implementation of on-site water reuse systems is hindered by the limited ability to ensure the level of treatment and protection of human health during operation. In this study, we tested the ability of five commercially available online sensors (free chlorine (FC), oxidation-reduction potential (ORP), pH, turbidity, UV absorbance at 254 nm) to predict the microbial water quality in membrane bioreactors followed by chlorination using logistic regression-based and mechanism-based models. The microbial water quality was assessed in terms of removal of enteric bacteria from the wastewater, removal of enteric viruses, and regrowth of bacteria in the treated water. We found that FC and ORP alone could predict the microbial water quality well, with ORP-based models generally performing better. We further observed that prediction accuracy did not increase when data from multiple sensors were integrated. We propose a methodology to link online sensor measurements to risk-based water quality targets, providing operation setpoints protective of human health for specific combinations of wastewaters and reuse applications. For instance, we recommend a minimum ORP of 705 mV to ensure a virus log-removal of 5, and an ORP of 765 mV for a log-removal of 6. These setpoints were selected to ensure that the percentage of events where the water is predicted to meet the quality target but it does not remains below 5%. Such a systematic approach to set sensor setpoints could be used in the development of water reuse guidelines and regulations that aim to cover a range of reuse applications with differential risks to human health.

Adaptive soft sensing of river flow prediction for wastewater treatment operation and risk management

Many wastewater utilities have discharge permits directly tied with the receiving river flow, so it is critical to have accurate prediction of the hydraulic throughput to ensure safe operation and environment protection. Current empirical knowledge-based operation faces many challenges, so in this study we developed and assessed daily-adaptive, probabilistic soft sensor prediction models to forecast the next month's average receiving river flowrate and guide the utility operations. By comparing 11 machine-learning methods, extra trees regression exhibits desired deterministic prediction accuracy at day 0 (overall accuracy index: 3.9 × 10^-3 1/cms²) (cms: cubic meter per second), which also increases steadily over the course of the month (e.g., MAPE and RMSE decrease from 41.46% and 23.31 cms to 3.31% and 2.81 cms, respectively). The overall classification accuracy of three river flow classes reaches 0.79 at the beginning and increases to about 0.97 over the course of the predicted month. To manage the uncertainty caused by potential false negative classification as overestimations, a probabilistic assessment on the predictions based on 95% lower PI is developed and successfully reduces the false negative classification from 17% to nearly zero with a slight sacrifice of overall classification accuracy.