Novel approach for AI-based N₂O emission reduction in biological wastewater treatment relying on genetic algorithms and neural networks Open Access

Global efforts to minimize greenhouse gas emissions require a significant reduction of CO₂-equivalent (CO₂e) emissions in all sectors. The wastewater sector can also essentially contribute, e.g. by minimizing nitrous oxide (N₂O) emissions during biological wastewater treatment (high global warming potential of 273 g CO₂e/g N₂O, IPCC (2021)). Within the scope of biological nitrogen removal, the formation and release of N₂O are subject to high temporal dynamics and are influenced by a variety of different factors (Zhou et al. 2008; Chandran et al. 2011; Schneider et al. 2013; Ribera-Guardia et al. 2014). However, N₂O emissions can be reduced to a minimum through control strategies that are individually tailored to the biological treatment process (e.g. Parravicini et al. 2015). These strategies usually rely on on-site measurements and predefined relationships between measurable parameters and control variables (e.g. Duan et al. 2020; Freyschmidt & Beier 2022). Recent control strategies focus on optimized aeration strategies (e.g. Duan et al. 2020; Freyschmidt & Beier 2022), enhanced denitrification (e.g. by dosing carbon, Peng et al. 2017), or optimized feeding (e.g. Valkova et al. 2021).

Not all parameters that influence N₂O formation and emission can be measured directly, or much effort is involved to obtain the required data due to significant spatial and temporal variations. This limits the potential of measurement-based plant control from a practical point of view. In contrast, a calibrated digital model (e.g. ASM3/1_N2O_ISAH, Beier et al. 2021) running parallel to plant operation (‘digital twin’) ensures that more parameters (also those that are not measured) can be integrated into a control algorithm (Seggelke et al. 2005; Matheri et al. 2022). Model-based control thus enables the implementation of extended control strategies for more complex objectives. Nevertheless, the main disadvantage remains that these models are still based on fixed control relationships.

However, the optimum operational settings for minimizing N₂O emissions can also be determined individually for an operating situation based on predictions, also considering legal effluent restrictions and energy demand. In this approach, various sets of operational settings are tested with the model, and the resulting N₂O emission is predicted. Finally, the dataset resulting in the lowest emission is chosen. It is no longer necessary to define fixed individual control relationships for specific targets; instead, the system is optimized from a holistic perspective.

Choosing the optimum control settings for a specific operating situation to minimize N₂O emissions can be automated with the help of intelligent, nature-inspired optimization algorithms. Unlike conventional derivation-based algorithms, nature-inspired algorithms like the genetic algorithm do not make any demands on the steadiness and differentiability of the optimization function, enabling the integration of various models (Nissen 1997).

For a prediction-based control approach, a large number of predictions may be necessary for a single optimization step. For that reason, the direct use of mechanistic models is currently considered unfeasible due to their high complexity and long simulation times. There is still a lack of quick-response tools that can be integrated into the operating practice.

As an alternative to mechanistic modeling, neural networks have become a focus of research in recent years (e.g. Hwangbo et al. 2021; Li et al. 2022). Neural networks are counted among the methods of artificial intelligence (AI) and consist of multiple artificial neurons, which transfer several inputs to an output based on the weighted sum of the inputs, a bias, and an activation function (Backhaus et al. 2015). By automatically changing the network parameters (weights, biases) during the network training, neural networks are capable of iteratively approximating complex mathematical (non-probabilistic) relationships between input and output data (Backhaus et al. 2015). The simplest type of neural network is a feedforward network, in which information is only passed in one direction (Backhaus et al. 2015). Compared to mechanistic models, neural networks require fewer input parameters for the prediction of N₂O emissions and are much faster due to their mathematically simpler structure. This makes neural networks ideal for dynamic process optimization. However, extensive training data (several annual series, including the parameters required to describe the operating state, the parameters that influence N₂O emissions, and the resulting N₂O emissions) is required to train the networks. In practice, such data sets are often not available.

This paper presents a novel method for intelligent plant control, aiming for the minimization of operational N₂O emissions during biological wastewater treatment. The approach includes three main components. First, operational settings associated with minimum N₂O emissions are identified by relying on a genetic algorithm. Second, aiming for short simulation times, neural networks were used as part of the optimization function to compute the fitness value of each candidate solution. Third, a classification algorithm was implemented to evaluate the reliability of the proposed control strategy.

The pilot plant (see Figure 1) was equipped with online sensors for temperature, pH, and oxygen concentration. In addition, the influent and effluent composition (Q, NH₄–N, NO₂–N, NO₃–N, alkalinity) were analytically determined every week. During two measurement campaigns, lasting 3 and 5 days (stable operating conditions, small temporal changes in N₂O formation), the N₂O concentration in the liquid phase was measured; in one of these measurement campaigns, the N₂O concentration in the reactor's off-gas and the off-gas volume flow were also determined to obtain data for the calibration of the gas transfer model.

A mechanistic model of the pilot plant was set up, using the expanded ASM model ASM3/1_N2O_ISAH (Beier et al. 2021). For calibration, seven data sets (influent and effluent composition (see Supplementary material, Table S1), O₂ concentration, wastewater temperature, pH value) of a specific operating phase (duration: 49 days) were employed. The model was calibrated for nitrogen conversion as well as N₂O formation and emission. During the investigation period, the commissioning of the system had already been completed, and a stable nitritation limited by the alkalinity was achieved. The biofilm thickness still increased slightly. More information on the model calibration can be found in Freyschmidt & Beier (2022). After completing the calibration process, nitrogen conversion as well as N₂O formation and emission could be depicted with an accuracy being sufficient for the scope of this study (see Supplementary material, Figures S1 and S2). As described in Freyschmidt & Beier (2022), N₂O formation can be traced back to the activity of autotrophic bacteria during the aerated phases inside the biofilm. However, also heterotrophic bacteria contribute to the N₂O formation, due to a strong inhibition of the N₂O reductase by nitrous acid (HNO₂). The gas transfer rate (and therefore the emission) is limited by the diffusion rate from the biofilm to the bulk phase.

As a second example, the methodology for emission reduction was applied to a model of the large-scale WWTP Kralingseveer. This example does not address the implementation of the developed approach during practical operation in real time. Instead, the approach was tested on historical data provided by Daelman et al. (2015). Only complete datasets were employed, including N₂O emissions, operational settings as well as influent and effluent composition (in total 55 datasets, covering a time span of 14.5 months). The biological stage of WWTP Kralingseveer consists of a plug flow reactor followed by two carousel reactors (flow scheme in Supplementary material Figure S3, more information in Daelman et al. (2015)). In this work, only the results of applying AI-based control to the model of WWTP Kralingseveer are summarized and discussed. More information on the setup of the algorithms and models can be found in the Supplementary material.

The same biological model was used to create the mechanistic model as for the pilot-scale example (expanded ASM model ASM3/1_N2O_ISAH; Beier et al. 2021). However, the calculation of the N₂O formation factor as a function of the nitrite concentration was adapted to mainstream conditions (due to a lack of adaption, much lower nitrite concentrations can already induce N₂O formation, e.g. Daelman et al. (2015) and Kuokkanen et al. (2021)).

For implementing the genetic algorithm and for setting up the neural networks, MATLAB R2023a (especially Optimization Toolbox, Global Optimization Toolbox, Deep Learning Toolbox, and Statistics and Machine Learning Toolbox) was employed.

The operational settings (O₂ target concentration, on/off times of intermittent aeration) of the investigated pilot plant were optimized using a genetic algorithm, aiming to minimize the target value ‘CO₂e emissions’. Only multiples of 0.5 were accepted for the O₂ concentration. The intermittent aeration on/off times were randomly selected from three predefined scenarios (on/off = 15 min/15 min, 15 min/30 min, 30 min/30 min). With all combinations, a stable suppression of nitratation could be achieved during the pilot plant operation.

In this case, an interval of 1 week would be the obvious choice for calculating the CO₂e emissions resulting from a specific set of operational settings, as the influent composition of the pilot plant changes weekly. However, since pre-investigations have shown that the plant performance stabilized approximately 3 days after the influent was changed, an interval of 3 days was selected.

In addition to the primary goal of minimizing CO₂e emissions, further boundary conditions were defined to ensure a stable nitritation (NH₄–N conversion ≥20%, effluent alkalinity <2 mmol/L, ratio of NO₃–N effluent concentration and NO₂–N effluent concentration <2%). It has to be emphasized that the mean NH₄–N conversion during plant operation was 24%, limited by alkalinity. The fulfillment of the boundary conditions is verified by relying on a prediction of the effluent concentrations of NH₄–N, NO₃–N, and NO₂–N as well as the effluent alkalinity.

Further information on the parameterization of the genetic algorithm can be found in the Supplementary material.

As indicated before, tools are needed to rapidly predict the N₂O emission as well as the effluent concentrations of NH₄–N, NO₃–N, and NO₂–N and effluent alkalinity, depending on the operational settings. However, the extent of the available operating data was not sufficient for the training of reliable neural networks for prediction. For that reason, a novel approach to transfer a mechanistic model into multiple neural networks was developed in this work. Mechanistic models can already be calibrated with data from shorter measurement campaigns and are based on mathematical relationships so that the network training can benefit from existing process understanding that has already been converted into a mathematical form. The network can be regarded as an intermediate memory for mechanistic simulation results; however, reliable network outputs can also be calculated for input data that was not considered in exactly the same form during the training phase. Similar approaches were described by Mehrani et al. (2022) and Li et al. (2022). The authors employed a mechanistic model to increase the resolution of the operating data of an (experimental) plant; the datasets generated in this way were also employed to train neural networks.

In the first step, datasets for the network training were generated with the aid of the mechanistic model of the pilot plant and random model input parameters (influent volume flow, influent NH₄–N concentration, influent alkalinity, temperature, pH value, O₂ target concentration, and on/off times of intermittent aeration). The variation of the input parameters followed defined probability distributions (normal distribution for influent volume flow and composition and equal distribution for all other parameters), ensuring that also rarely occurring operational situations can be included in the network training with sufficient frequency. The range of the operational settings was limited as described above (multiples of 0.5 for the O₂ concentration, three predefined scenarios for intermittent aeration).

Subsequently, the resulting N₂O emissions for each input data set as well as the additional parameters that are needed to evaluate the boundary conditions (effluent concentrations of NH₄–N, NO₃–N, and NO₂–N and effluent alkalinity) were determined using the mechanistic model. The first simulation (first dataset) is carried out from a steady state; all further simulations are carried out from the resulting state of the previous simulation. By varying the starting states, it is taken into account that the N₂O emission depends not only on the current influent composition and the operational settings but also on the previous operating situations.

Finally, 20,000 datasets were generated via mechanistic modeling, including the influent volume flow and composition (NH₄–N concentration, alkalinity), the wastewater temperature and the pH value, the operational settings (O₂ target concentration, on/off times of intermittent aeration), the wastewater composition (concentrations of NH₄–N, NO₃–N, and NO₂–N and alkalinity) in the reactor before the simulation starts, the resulting N₂O emission, the resulting effluent concentrations of NH₄–N, NO₃–N, and NO₂–N and the effluent alkalinity.

For the setup of the neural feedforward networks to predict the N₂O emission as well as the effluent concentrations of NH₄–N, NO₃–N, and NO₂–N and the effluent alkalinity, input parameters were chosen from the data included in the generated training datasets according to Table 1. A total of 456 datasets resulting in an N₂O emission <0 were neglected. The networks are initialized with an input layer and two hidden layers; the number of neurons per layer was determined iteratively (see Table 2). The Levenberg–Marquardt algorithm was used for network training.

The reliability of the proposed operational settings is mainly driven by the quality of the N₂O emission prediction. To avoid counterproductive optimization (emission increase after applying the proposed control strategy), the operating situations in which the predicted N₂O emissions are not reliable have to be identified.

In this work, 60 neural feedforward networks (2 layers, 20 neurons + 10 neurons) are trained to classify the predicted N₂O emission. Only reliable and unreliable results are distinguished. If at least 45 networks rate the predicted N₂O emission as reliable, the prediction is considered reliable (‘bootstrap aggregating’ or ‘bagging’ according to Breiman 1996).

The training datasets generated with the mechanistic model are also used to train the classification networks. In order for the algorithm to evaluate the network performance during the training, each training dataset must be classified. For this purpose, the 95% percentile of the absolute deviation between the N₂O emission calculated with the mechanistic model and the one predicted with the neural network is employed.

For the training of the classification networks, all parameters that have already been used for the training of the prediction networks (19,544 datasets) as well as the predicted N₂O emission were used as input parameters. Moreover, parameters to evaluate the stability of the N₂O emission prediction, which are determined as percentage deviations between the N₂O emission calculated with the actual input parameters and the N₂O emissions calculated with input parameters increased or decreased by 1%, were integrated into the network training. Finally, the first principal component of the wastewater composition in the reactor before and after the simulation of the respective dataset (calculated with additional neural networks, R² = 1,000 and 0.999) was employed for network training.

For each network to be trained, 75% of the datasets for which the predicted N₂O emission was classified as unreliable are randomly selected for the training process. In addition, triple the amount of datasets with a predicted N₂O emission classified as reliable is selected (also randomly). Since misclassification of an unreliable N₂O emission is more critical due to a potentially counterproductive optimization, incorrect classifications are weighted differently in the performance evaluation during the network training (factor 10 for incorrect classification as reliable).

The performance of the network for predicting N₂O emissions was additionally evaluated with the N₂O emissions measured during the measurement campaign. The neural network predicts an average N₂O emission of 1.86 g N₂O-N/day over the duration of the measurement campaign (duration: 3 days). In situ, an N₂O emission of 1.95 g N₂O-N/day was measured. The deviation between calculated and measured emissions is approximately 5%.

In summary, the mechanistic model was successfully transferred to several neural networks so that a fast prediction of N₂O emissions and other target parameters can be made based on measured values only. So, a considerable reduction in model complexity (fewer input parameters and simpler equations) can be achieved while maintaining nearly the same model accuracy.

Independent of the integration into an optimization algorithm, the approach of transferring a mechanistic model into neural prediction networks has a high potential for practical application. A comparable accuracy can be achieved with a model that is much easier to handle. In addition, the number of required input parameters can be reduced, which simplifies operational use. If sufficient data are available for the calibration of a mechanistic model, the approach can be transferred to any wastewater treatment parameters; the prediction of N₂O emissions represents one of the most challenging cases here due to the large number of influencing factors. It has to be pointed out that the network training must be renewed in the case of changing operating conditions or wastewater composition, which are not covered by the range of the training data.

The optimal operational settings were determined for the seven datasets already used for the calibration of the mechanistic model. The results show that the algorithm identifies intermittent aeration with an aerated and non-aerated phase of 15 min each as the optimal aeration strategy for all operating situations considered (see Table 3). The O₂ concentration is selected so that the required NH₄–N conversion is achieved or the maximum alkalinity in the effluent is just not exceeded. Higher O₂ target concentrations are therefore recommended for higher NH₄–N influent loads or higher alkalinity in the influent.

N₂O mitigation is achieved by ensuring a sufficient oxygen supply in the aerated phase, reducing the autotrophic N₂O formation. Due to the short aerated interval, less N₂O accumulates in the biofilm, so the diffusion rate is reduced by a lower concentration gradient. The accumulated N₂O is partly converted by heterotrophic bacteria in the anoxic phase. In this context, short aeration intervals are also advantageous since less nitrite accumulates. Thus, the concentration of inhibiting HNO₂ is also lower. Similar operational strategies for mitigating N₂O emissions were identified by Peng et al. (2017). The authors also achieved a reduction of N₂O emissions with short intermittent aeration intervals.

The CO₂e emissions that would result from applying the proposed control parameters are in all cases lower than the emissions recorded in the measurement campaign (carried out under the conditions of dataset 5, 3.06 kg CO₂e over the investigated interval of 3 days). However, the N₂O emissions of the non-optimized system were not measured for all datasets, so the total potential CO₂e emission reduction cannot be quantified.

For dataset 1, the mechanistic modeling shows that with the ‘15/30’ aeration strategy, an alkalinity of 2 mmol/L is exceeded in the effluent regardless of the target O₂ concentration (violation of boundary condition 2). This is also observed with the other aeration strategies if the O₂ target concentration falls below 1.5 mg/L. In these cases, nitrogen conversion is no longer limited by the alkalinity but by the O₂ supply. With the ‘15/15’ aeration strategy, the demanded suppression of nitrate formation cannot be achieved at higher O₂ target concentrations (≥4 mg/L) (violation of boundary condition 3).

Both the optimization algorithm and the mechanistic model achieve the lowest emissions with the ‘15/15’ aeration strategy. The optimization algorithm proposes a target O₂ concentration of 3 mg/L; the mechanistic model results in the lowest emissions with a target O₂ concentration of 3.5 mg/L. This deviation is due to minor inaccuracies in the prediction of the NO₃–N effluent concentration by the corresponding neural network, which leads to a violation of boundary condition 3. Thus, the optimization algorithm only finds the potentially second-best solution (depending on the accuracy of the mechanistic model); however, the corresponding CO₂e emission only exceeds the CO₂e emission at optimal operating settings by 4%.

The presented results confirm that the developed method for intelligent plant control can support the reduction/prevention of operational greenhouse gas emissions during wastewater treatment. The deviations between the solutions determined with the algorithm and the mechanistic model are considered to be negligible, as the mechanistic model is also associated with model uncertainties (e.g. regarding the integration of influencing parameters on N₂O formation). Instead, it can be emphasized that similar control parameters were identified with both methods, which confirms the practical applicability of the proposed approach. It also has to be stressed that the optimization algorithm produces a result much faster and works independently. While the intelligent optimization method determines optimal operational settings within a few minutes (<2 min), the application of the mechanistic model needs much more time (approximately 15 min), since all possible combinations have to be investigated. In addition, it must be manually checked for which control parameter combinations boundary conditions are violated.

The presented approach was employed to exemplarily determine the potential for emission reduction for the large-scale WWTP Kralingseveer (data from Daelman et al. 2015). The developed methodology was applied as described before (setup and calibration of a mechanistic model using the biological model ASM3/1_N2O_ISAH, training of neural networks, preparation of the algorithm for optimization, and result evaluation). Due to the complexity of the investigated plant, not only measured parameters but also parameters that were calculated with a digital twin during operation were employed to set up the neural networks for prediction and result evaluation. Altogether, 94 network inputs were used, including wastewater composition, operational parameters, and biomass concentrations at different positions of the biological treatment stage (performance of trained prediction networks and result evaluation algorithm in Supplementary material). Maximum values for NH₄–N (4 mg N/L), NO₂–N (0.6 mg N/L), and inorganic N (11 mg N/L) effluent concentrations were defined as boundary conditions for optimization.

The total potential emission reduction was determined by comparing the emissions that occurred during operation (measurement data available) with those of the optimized system. During the optimization process, the target oxygen concentrations in all aerated zones as well as the recirculation rates of both stages were varied. It has to be emphasized that the plant was not operated with the aim of minimizing N₂O emissions.

Figure 6 show the measured emissions as well as the emissions predicted by the algorithm and calculated by the mechanistic model (control of the algorithm). Using intelligent plant control, a potential emission reduction of 43% or 1,588 t CO₂e/a seems possible. The modeled N₂O emissions decreased by more than 50% (N₂O emission factor: 3.9% of oxidized nitrogen before optimization and 1.9% after optimization). The reduction was mainly achieved by establishing low oxygen concentrations in the aerated reactor zones. The oxygen supply was limited to the amount required for full nitrification. So, additional simultaneous denitrification capacities could be activated. Nevertheless, the achieved nitrogen elimination was comparable to the performance of the non-optimized system.

The work presented here confirms the high potential of intelligent plant optimization for solving complex control tasks in biological wastewater treatment, like reducing the operational CO₂e emissions. Intelligent algorithms find highly effective control strategies in a very short time for the majority of operating situations. However, this requires the availability of accurate prediction tools. The key findings are as follows:

• An accurate prediction of N₂O emissions and other target parameters during biological wastewater treatment was achieved by transferring a calibrated mechanistic model to several neural networks (one network for each parameter to be predicted).

• The use of neural networks significantly increased the simulation speed compared to the mechanistic model. The prediction quality of the mechanistic model and the neural networks was shown to be comparable (R² > 0.90).

• Neural networks are reliable in their predictions if the operating situations for which the networks are to be used are represented with sufficient frequency in the training datasets. With the method presented here, it is possible to identify the majority of operating situations for which reliable emission prediction is not possible.

• The exemplary application of the optimization algorithm showed the high impact of the thoughtful definition of the boundary conditions (e.g. achieving a certain effluent quality) on the practical suitability of the recommended operating settings. If an operational target is not defined, it is not taken into account during optimization.

• By applying the proposed control strategies, a potential reduction in N₂O emissions of more than 50% was achieved in the large-scale example investigated.

Overall, it was proven that the proposed method enables a new type of operational support for wastewater treatment based on neural networks. In the future, intelligent algorithms will enable even further automation of wastewater treatment plant operations and will contribute to more sustainable, lower-emission operations.

The content presented in this publication is part of a PhD thesis submitted and accepted at the Leibniz University Hannover (Freyschmidt, Arne: ‘N₂O-Emissionsreduzierung durch KI-unterstützte Regelung der biologischen Abwasserreinigung’ – ‘N₂O emission reduction by AI-supported control of biological wastewater treatment’, 2024).

We thank the Ministry for Science and Culture of Lower Saxony and the VolkswagenStiftung for the funding within the framework of the 'Zukunftslabor Wasser', the German Federal Ministry of Education and Research for the funding within the framework of the MiNzE project (`Minimization of the CO2 footprint by adapted process development in process water treatment - testing of the MiNzE process in an immersed fixed bed', FKZ: 02WQ1482B), and we gratefully acknowledge the financial support of the German Research Foundation (DFG) in procuring the experimental and measurement set-up for the in-depth measurements of biotechnological processes. The publication of this article was funded by the Open Access Fund of Leibniz University Hannover.

A. F. conceptualized the work, developed the methodology, contributed in programming and modeling, writing, editing. S. K. conceptualized the work and reviewed the article.

All relevant data are available from an online repository or repositories: https://data.uni-hannover.de/de/dataset/data-freyschmidt-et-al-2025.

The authors declare there is no conflict.