Modelling the adoption of combined sewer overflow surveillance in Switzerland Open Access

What are the most effective policy instruments to foster the adoption of monitoring in combined sewer overflows (CSOs)?

CSOs release untreated wastewater when heavy rainfall causes runoff volumes to exceed the hydraulic limits of sewer systems. CSOs can be an often-overlooked yet significant source of pollutants, contributing levels of microplastics, particulates, faecal bacteria, and nutrients comparable to those from wastewater treatment plants (WWTPs) (Milieu 2016; Pistocchi & Dorati 2018).

While substantial investments have been made to upgrade WWTPs (Eawag 2015), utilities may find it highly cost-effective to reduce pollution by improving CSO performance (Dirckx et al. 2011; Bachmann-Machnik et al. 2021) or optimizing the entire wastewater management system (Meng et al. 2016). In regions where CSO monitoring exists, data reveal that overflows occur with concerning frequency, indicating substantial pollution levels (Guardian 2021; Giakoumis & Voulvoulis 2023).

This observation aligns with recent updates to the Urban Wastewater Treatment Directive (UWWTD), which now requires CSO monitoring and reporting (Rieckermann et al. 2021; European Commission 2024). Such regulations mirror approaches, such as the nine minimum technology-based controls set by the US Environmental Protection Agency in 1994, which aim to manage CSOs through operational adjustments rather than extensive new infrastructure (EPA 1994). Despite monitoring technologies existing for decades – and advanced solutions like real-time emission controls discussed 35 years ago (Schilling 1989) – CSO monitoring remains unevenly adopted. This raises critical questions about the motivations and barriers for cities and utilities to implement CSO monitoring on a broader scale.

In the United Kingdom, top-down regulation has proven effective; following a parliamentary mandate for CSO monitoring in 2013, action proceeded quickly within the privatized sector, dominated by a few major utilities. Since 2020, numerous datasets have been available to the public, such as on the government's homepage (Defra 2024).

In Germany, however, top-down regulation has had mixed results. North Rhine-Westphalia, for instance, initially introduced a self-compliance approach, which was recently reinforced with a formal state regulation to mandate CSO monitoring (Hoppe et al. 2018). The limited impact of the initial regulation underscores challenges in Germany's more fragmented water industry, which includes both large regional associations and numerous small, publicly owned utilities. Similarly, in Baden-Württemberg, a state regulation on CSO monitoring has existed for several years (EKVO 2001), yet compliance data have often been found inaccurate (Dittmer et al. 2015). This has been attributed to factors like inconsistent sensor maintenance, limited skills among local operators, and data aggregation errors within Supervisory Control and Data Acquisition (SCADA) systems. To address these challenges, the regional German water association in Baden-Württemberg launched a ‘bottom-up’ community-driven program to improve stakeholder education and promote knowledge-sharing networks (DWA 2025).

Switzerland's current legislation does not explicitly require the monitoring of CSOs (Rieckermann et al. 2017). Yet, some operators, as well as cantonal regulators, interpret the legal requirements¹ to ensure the ‘economic operation of public sewers’ by utilizing sensors. Also, installing sensors to gather data about CSO performance is considered a ‘reasonable measure’ to minimize the discharge of harmful substances into water bodies by some but not all operators.

These cases illustrate the ongoing uncertainty about the most effective policy instruments to foster the adoption of CSO monitoring as a best practice to reduce overflow emissions, particularly in regions with diverse organisational structures, resource constraints, or limited technical expertise (Manny et al. 2021, 2022). This interplay between socio-technological systems can be investigated with ABMs.

ABMs are computational simulations that represent interactions between autonomous agents within a shared environment to study complex system behaviours (Niazi & Hussain 2011). While they are used in a wide range of applications such as epidemic spread (Marini et al. 2020) or traffic simulation (Saprykin et al. 2019), this paper focuses on ABMs that address non-technical barriers to support policy decisions.

In urban drainage, ABMs have been increasingly applied in infrastructure planning, for example, to generate virtual sewer networks (Urich & Rauch 2014) and flood-risk management (Zhuo & Han 2020), but applications in policy decisions are missing. Policy decisions in drinking water have been modelled with ABMs, where the behaviour of stakeholders in the drinking water supply of the city of Zurich (Tillman 2001) considered involving directly in the development and validation of a ‘rule-catalogue’. This rule-catalogue then was at the core of describing the building of water supply infrastructure in the model, which is a very attractive concept to capture the main behaviour of involved agents.

More recently, ABMs were applied in the domain of technology diffusion, where important processes concern the exchange of opinion between agents. In this regard, the Theory of Planned Behaviour (TPB) (Ajzen 1991) is a suitable cognitive model to represent the agents' behaviour with a well-established psychological theory. For example, Schwarz & Ernst (2009) used TPB to study the diffusion of water-saving innovations such as dual-flush toilets, founding their agent-based model (ABM) empirically on a survey conducted for this purpose. Anebagilu et al. (2021) successfully apply ABM and TPB for policy support in the context of protecting water bodies through vegetative filter strips. Rai & Robinson (2015) apply a similar methodology to the adoption of residential solar photovoltaic and model the agents' interactions through the bounded confidence opinion dynamics (BCOD) model (Hegselmann & Krause 2002; Bernardo et al. 2024). However, the application of the ABM method to investigate technology diffusion for CSO monitoring is lacking.

In this study, we, therefore, suggest a prototype ABM for the adoption of monitoring technology in CSOs in Switzerland, developing scenarios that are based on different policy instruments such as regulations and technology improvements. To this aim, we combine the foundational method of TPB with an adapted ‘rule-catalogue’ (Tillman 2001) to root our assumptions in the reality of Swiss urban wastewater management through expert elicitation. Furthermore, we integrate BCOD to describe how stakeholders exchange opinions on a social network.

This is novel in three distinct ways. First, we thoroughly analyse the roles of different stakeholders in Swiss wastewater management and their impact on the adoption of CSO monitoring. Second, we apply the ABM including TPB and BCOD to the case of monitoring technology adoption in CSOs in Switzerland and assess its usefulness. Third, we offer an open implementation of our prototype (Derwort 2021), making it available for further development, extension, or adaptation to other contexts – such as supporting the rollout of the revised UWWTD and its new CSO monitoring requirements across Europe. This paper builds on the initial work by Derwort (2021), expanding the use of ABM for exploratory analysis while improving the reusability of the code base. Also, we integrate recent developments, particularly in socio-technical network analysis, open governance data in the UK and updates related to the UWWTD recast. We also improved the critical assessment of the prototype's model outputs.

In the remainder of the paper, we describe our methodology for developing the prototype. The ‘Results’ section presents findings from various scenarios simulating policy instruments, with sensitivity analyses to examine the prototype's behaviour. In the discussion, we evaluate the implications of these findings, the limitations, and the challenges around data availability.

In this article, we define the adoption of CSO monitoring practices as ‘sensor adoption’, which corresponds to a functioning, calibrated sensor being installed within a CSO tank and data are continuously collected, either on a local datalogger on-site or off-site in a SCADA system. We consider this a reasonable simplification for the ABM prototype, since it is the first necessary step towards good CSO monitoring practices. The primary output of the prototype is the time series of WWTPs and municipalities that adopt sensors over the simulation period.

The flexible nature of ABMs, allowing them to be tailored to specific contexts, is at the same time a challenge for their calibration and validation, which is also due to the complexity of the systems at hand. It lies in the nature of a complex system that it cannot be captured well under most definitions of complexity which however should not hinder our efforts to do so (Bankes 2002).

There is no one-size-fits-all method for validating ABMs, as it depends on the model and the available data for comparison. In ABM, identifying quantitative parameters using a likelihood approach is uncommon, among other things because limited data can lead to insufficient constraints of parameters and model structure (Srikrishnan & Keller 2021). Other issues concern structural uncertainties, long runtimes and the need for multiple observation levels (Fehler et al. 2006).

In our case, empirical validation is challenging, because historical data, i.e. time-series of sensor installation in certain areas, does not exist for Switzerland. This lack of temporal data also renders a calibration of the opinion dynamics impossible. Thus, we refrain from fitting the model in this regard, since it would potentially obscure interesting and valuable behaviours generated by the simulation.

While relying on TPB as a well-established theory for the agents' decision-making, expert consultation played a key role in validating the model's assumptions. The ‘rule-catalogue’, adapted from Tillman et al. (1999; Tillman 2001), was originally designed to describe agent behaviour through if-then rules intended to be validated by domain experts. Although time constraints prevented extensive expert meetings, this approach enabled us to ground the model's assumptions in reality. Experts provided estimates for key model parameters, such as communication schedules.

To characterize the uncertainties in the model and evaluate model behaviour, a sensitivity analysis (SA) is performed, as described later in this section. Furthermore, numerical experiments are performed to evaluate the model behaviour in different configurations and provide a deeper understanding of the dynamics at play in the system.

The ABM that we propose incorporates the organizational and institutional situation in Switzerland and is in its current implementation entirely designed and applicable to this local scope (Derwort 2021).

In Switzerland, the regulation and execution of urban water management competencies are situated at the level of the sub-states (so-called ‘cantons’). Sub-state authorities are public administrations with the obligation to monitor compliance with water protection regulations, for example. Operational competencies such as the daily operation of sewer systems and WWTPs are in most cases fulfilled by municipalities or wastewater associations (Luís-Manso 2005). This public service provision is complemented by further important stakeholders, such as consulting engineers and industry firms.

Based on the simplification of stakeholders in Swiss urban water management, we initialize the four different agent types in the following way. Each municipality agent is generated based on geodata available for Switzerland (see Table 1). The municipality represents an aggregation of multiple individual stakeholders such as municipal engineer, municipal council, municipal president. For simplification, we assume that each WWTP has one WWTP operator and assign a corresponding agent for each of the 720 WWTPs in our database (Maurer & Herlyn 2006). Engineers act as important stakeholders in terms of infrastructure planning and service provision. ‘Engineer’ refers to large engineering consulting firms that operate across both geographical and language regions in Switzerland, for which a number of 5 agents was estimated by experts. In Switzerland, 26 cantonal (sub-state) authorities are responsible for law enforcement and compliance assessment. The ABM includes in total 26 authority agents, one for each canton (sub-state). For details see Derwort (2021, section 2.3.5).

In our ABM, we chose to exclude stakeholders at the national level, as well as professional associations and civil society, to maintain a focused and parsimonious model. Instead, we emphasize regional (cantonal) authorities, who hold delegated responsibilities from the national government (Manny 2017) and are more directly involved in regular interactions with municipalities and WWTP operators. This focus aligns with the monthly timestep of our simulation.

To capture the influence of national-level policymaking, we included the ‘Regulation’ scenario in the model. While we excluded professional associations such as the Swiss Wastewater Association (VSA), we acknowledged their important role in cross-regional knowledge exchange. This function is represented through the ‘Engineer’ agents, whose employees are actively involved in developing VSA guidelines and participating in training events.

As mentioned in the introduction, TPB (Ajzen 1991) has gained popularity as a cognitive model for agents in ABMs (Schwarz & Ernst 2009; Rai & Robinson 2015; Anebagilu et al. 2021), especially in the context of sustainability and environmental protection. We consider it a suitable model for the decision-making of agents in our case of sensor adoption in CSOs.

To represent the interactions between agents, our ABM draws on the relative agreement (RA) algorithm (Deffuant et al. 2002) which has been used in social simulations and includes the theoretical concept of BCOD (Hegselmann & Krause 2002). It assumes that agents influence and are influenced by other agents whose opinion lies within a certain interval (Manny et al. 2022). A flowchart and detailed information on the RA algorithm are given in Derwort (2021, Appendix E). To characterize the range of opinions differing from its own which an agent is ready to consider, the uncertainty u is introduced. It results directly from the formula u= |0.5 – sia|, assuming that extreme opinions coincide with smaller uncertainty, as shown on the bottom of Figure 4.

The RA algorithm includes a time scheduling, i.e. defines interactions at moments in time. For our ABM, we chose 1-month time intervals, which we deem reasonable time steps for inter-organizational communication. Communication schedules allow us to specify which agent types how often with others, e.g. every month or only once in a year (Derwort 2021, section 2.3.3).

To estimate the agents' attitude towards sensor installation (sia) and their perception of the difficulty of this endeavour (pbc), we combine different sources of knowledge (Table 1). Given that this ABM is only a prototype, we believe that using existing data are adequate for the purpose and the level of complexity needed, such as response data from a survey not conducted after TPB methodology (Manny 2017).

For the WWTP operators and cantonal authorities, we assess the survey responses concerning know-how, peer influence, and vision for the future and obtain values of sia and pbc for our survey respondents. From the survey among the cantonal authorities, we also have an estimate of the share of WWTP operators in a canton that has sensors installed. This share defines how many of the WWTP operators are initialized with sensors already installed. The sia and pbc values of a survey respondent are now assigned to an agent based on whether it has a sensor or not. This distribution for these values is shown in Figure 4.

For the municipalities and engineers, however, comparable survey data are unavailable. Here, we rely on qualified guesses of the attitude (sia) of municipalities towards sensor installation. We asked domain experts to estimate the distributions from which we sample the values in a questionnaire specifically designed for that purpose. The sources of our estimates are listed in Table 1, while a detailed description is provided in Derwort (2021, Appendix C).

The ABM includes a social network structure to represent interactions between agents. However, as empirical social network data was unavailable, the social network was created based on several weak assumptions. Examples from the complete list of 10 assumptions (Derwort 2021, section 2.3.6) are:

• WWTP operators communicate with other WWTP operators whose catchment areas are adjacent to theirs.

• Municipalities communicate with the cantonal authority of the canton they belong to.

• Cantonal authorities communicate with other cantonal authorities of adjacent cantons.

• Each engineer communicates with all other engineers (complete network).

Besides these agent-specific assumptions, agents further communicate on a local level, as it is also assumed for the generation of social networks in other ABMs (Rai & Robinson 2015).

Using these assumptions as well as the geospatial datasets for canton, municipality, and WWTP catchment area boundaries (Derwort 2021, Appendix C), the social network structure is obtained. During the initialization of the ABM, the network is transformed into a bidirectional graph structure. This graph serves as the environment within which agents communicate.

As the aim of this work is to assess strategies that could support the adoption of sensor technology in CSOs, we construct the following three scenarios (see Table 2) representing different policy instruments, e.g. regulations or events for wastewater professionals (Manny 2017). A baseline scenario for comparison incorporates the best knowledge available and an estimation of reasonable parameters. For details see Derwort (2021, Appendix F).

Scenario 1 ‘No joint planning meeting’ serves to evaluate the effect of joint planning meetings between the engineer, cantonal authority, WWTP operator and municipality of a catchment. We believe that these meetings play an important role for the establishment of, e.g. CSO monitoring. As the planning meeting is incorporated in the baseline scenario, this policy scenario is implemented through the removal of the yearly meeting among the mentioned participants from the communication schedule.

Scenario 2 ‘Regulation’ represents a top-down policy in which authorities require sensor adoption by WWTP operators and municipalities – comparable to a national regulation mandating CSO monitoring, such as one issued by the ministry of environment. In the simulation, this scenario is implemented by gradually increasing each agent's perceived social pressure to comply, represented by the ‘subjective norm’ component of the sensor installation attitude.

This increase reflects the idea that, over time, agents experience growing expectations from regulatory authorities, peers, and the broader public. Examples of such real-world mechanisms include formal communication, compliance monitoring, reporting obligations, or public benchmarking – factors known to shape organizational behaviour. Since no empirical data currently exist to quantify how quickly or strongly this normative pressure builds, the model applies a fixed incremental increase in sia at each time step as a simplified representation.

Scenario 3 ‘Better technology’ assumes that technology evolving into being more user-friendly and affordable makes WWTP operators and municipalities perceive sensor installation in CSOs as easier. This policy scenario is implemented using a lower pbc-threshold over time to simulate. This scenario is motivated by recent developments in measurement technology, e.g. affordable and exact level sensors and wireless transmission technology.

A SA was conducted to identify the most significant parameters influencing the model's behaviour and outcomes. In this analysis, agent parameters were sampled from distributions and their means were varied systematically. A simple screening method, specifically the Morris method, was employed due to its computational efficiency and suitability for this early stage of prototype development, given its relatively low computational cost.

In addition to this SA, a non-exhaustive robustness analysis was performed through numerical experiments to assess the implications of the model's inherent structure. These experiments included modifications such as shifts in the timing of communications between agents, by using different communication schedules from consultations of different experts, and changes in the network structure. For example, the model tested a fully connected graph for cantonal authorities as opposed to a more localised network. The opinion dynamics were also explored, with varying assumptions about the uncertainty of opinions.

The ABM was implemented in Python and is composed of three main components. The first module is responsible for loading and analysing input data using the Pandas and Geopandas libraries. This module constructs the social network of agents and initialises their attributes, setting up the foundation for agent interactions.

The second module leverages the Mesa framework, a Python library designed for building ABMs, to run the dynamic simulations. This module handles the interactions between agents over time, enabling the simulation of different scenarios and agent behaviours.

The third module provides scripts to execute the model in various configurations, perform sensitivity analyses, and generate visualisations of the results. This includes plotting the outcomes of different simulation runs, which aids in interpreting the model's behaviour under varying conditions. Detailed instructions for running these simulations and analysing the results are provided in the ABM's documentation (Derwort 2021).

Additionally, the SA was also implemented in Python, using the Morris method from the SALib library, as described by Herman & Usher (2017). This method allowed for an efficient evaluation of the impact of different model parameters on the overall outcomes. Each run of the model took approximately 45 s on a computer equipped with an Intel i7-5500U CPU running at 2.40 GHz, with 8 GB of RAM and operating on Windows 10.

Among the scenarios evaluated, regulation appears to be the most effective policy instrument for promoting sensor adoption. The results suggest that regulation doubles the adoption rate from 20 to 40% within 4 years and by the end of the 10-year period, 60% of municipalities had installed sensors. In contrast, the impact of regulation or any other scenario on WWTP operators was minimal, likely due to the fact that many WWTP operators had already installed sensors or were resistant to adoption due to their initially low sia, which was derived from survey data, as can be seen on the left-hand side of Figure 4.

Other scenarios, including reduced communication through the GEP (joint planning meetings) and improvements in sensor technology, had smaller effects. The GEP scenario slightly slowed down the rate of adoption but did not significantly reduce the total number of sensors installed by the end of the simulation. Technological improvements, while making sensors more accessible (lowering pbc_threshold), led to only a slight increase in sensor adoption rates.

The numerical experiments generally produced results similar to the baseline scenario, as can be noted on the left of Figure 5. Experiments that tested different communication schedules (Experiments 1 and 2) showed no significant deviations in sensor adoption compared to the baseline. Similarly, Experiment 3, which introduced a different network structure for cantonal authorities, did not substantially alter the model's outcomes.

Experiment 4, however, which introduced randomly sampled opinion uncertainty to the opinion dynamics, revealed that the internal parameters governing the opinion exchange process were critical to the model's behaviour. The results of this experiment were more widely scattered compared to other scenarios.

The numerical experiments suggest that there is considerable uncertainty regarding the parameters of the opinion dynamics. Notably, the results of Experiment 4 show a wide variance compared to the other scenarios and numerical experiments. The numerical experiments being a non-exhaustive analysis, e.g. a further introduction of stochasticity regarding the interactions could have significant influence. Additionally, the SA highlights the sensitivity of opinion dynamics to the initial parameters, specifically those related to the opinion algorithm.

The RA Algorithm, which underpins the opinion dynamics in this study, has several limitations. The algorithm simplifies human behaviour by assuming static parameters, such as opinion uncertainty, which are difficult to measure in reality and may not remain constant over time. There is a general uncertainty regarding the formalization of processes. For example, the way that TPB is implemented in ABMs influences the outcome of models, regardless that the theory is identical (Muelder & Filatova 2018).

Our model does not account for external events like attention on polluted rivers, as observed during the Paris Olympics (Surfrider Foundation 2024) or the Oxford Boat Race (Castle 2024), or reports of UK citizens risking their health by swimming in polluted rivers to demand better CSO monitoring (Herman & Usher 2017). Media attention on such events usually strongly influences opinions, but this is not included within the current ABM.

The numeric results of the ABM need to be considered with caution. While the model suggests a slow and saturating dynamic in sensor adoption, evidence from other countries indicates that real-world outcomes might differ significantly. For instance, in the United Kingdom, sensor adoption was relatively quick, while in Germany, the process has been notably slower. It is important to emphasize that the absolute numbers generated by the model may not reflect reality since many parameters were chosen ad hoc.

As previously mentioned, an empirically grounded ABM requires substantially more data than available in our case, such as a survey among stakeholders according to TPB, more fine-grained data about wastewater organisations and their infrastructure for initialization of the agents as well as historical data about sensor installation.

ABM can play an important role in combining multiple sources of knowledge. Not only did the development of the ABM foster the collection of new data sets but also allowed us to approach the question at hand in a multidisciplinary manner through the integration of behavioural models such as TPB. By formalizing assumptions such that they can be implemented in a computer model, the context of the model is thoroughly investigated. The assumptions can then on one hand be easily validated by experts who deemed them at least partly true in the case of our model. On the other hand, their implications can be simulated and yield insights, even or especially when they produce unexpected results.

For instance, the simulation results counter-intuitively show a ‘saturation’ of sensor installation. This points out the possibility that the authorities and engineers may overestimate their influence on the WWTP operators and municipalities, especially since the SA also shows that cantonal authorities and engineers have relatively little influence over sensor adoption. As a result, policies that leverage existing positive attitudes among the right agents, such as transferring ownership of CSO tanks to WWTP operators, might be a more effective strategy than attempting to shift attitudes towards sensor installation among municipalities.

Another factor that may explain the outcome is that the ABM fails to consider the benefits of sensor installation. When we recognize that sensor adoption is just one part of a broader digitization strategy aimed at fully utilizing data, the advantages for stakeholders become clearer. The true promise of digitization in water management lies in achieving more with fewer resources, enhancing efficiency, and using data to optimize infrastructure operations (Schilling 1989; Fu et al. 2024). However, realizing this potential requires high-quality data, which remains a critical issue, as evidenced by experiences in Germany (Dittmer et al. 2015). Legislation and sensor adoption are not silver bullets; they necessitate a cultural shift, a data-literate workforce, and the right processes to maximize the value of data.

In reality, combinations of different policy instruments are often observed (Pakizer et al. 2020). ABMs could be employed to identify the most effective policy mix, such as promoting data literacy, without – or especially before – resorting to costly longitudinal studies that are often difficult to manage and do not guarantee a successful outcome (Sherwood 2005). Also, while the predictions of ABMs may not provide absolute accuracy, significant prediction uncertainties do not diminish their ability to offer effective decision support (Reichert & Borsuk 2005).

The current implementation of the ABM has limitations that highlight areas for improvement in future development. Addressing these would enhance model plausibility, as suggested by Tillman (2001). For instance, refining the selection of stakeholders could be beneficial. An individual agent in the model may represent multiple people – such as a cantonal authority with regional departments – while the opinion dynamics and the TPB focus on individual behaviours. Feedback from expert interviews also indicates a need to differentiate stakeholders within municipalities and wastewater associations into political, administrative, and technical domains. Additionally, including hardware and software providers of monitoring technologies could strengthen the model's alignment with practical scenarios. Although the prototype currently does not incorporate it, its structure allows easy adaptation to scenarios that combine different policy instruments. Further adaptation is needed to reflect varying organisational structures, such as those in the United Kingdom, because Swiss reality often differs significantly from other contexts (Blumensaat et al. 2012).

For CSO monitoring, it is essential to consider the broader value of data, not only for operational management but also for urban drainage planning by municipalities and compliance assessment by regulatory authorities. Assessing the financial and practical incentives of high-quality monitoring data would enable municipalities to compare the benefits of proactive data collection with those who disregard system performance data, maximizing both planning and operational capacities.

Follow-up work on the ABM could explore its potential to depict self-reinforcing dynamics such as the impact of examples of good monitoring practice. These would affect the agents' perceived behavioural control, e.g. in a scenario where agents would perceive monitoring technology as more accessible if they directly come into contact with best practice examples.

To better capture the perceived utility of monitoring data, the model could incorporate a utility function, accounting for factors like the impact of sensor adoption and the potential benefits of enhanced data (Scholten et al. 2017).

Also, recent theoretical development on standardized social network structures of actors in urban water management, i.e. ‘motifs’ (Manny et al. 2022; Manny 2023) could make it possible to directly consider empirical motifs in the ABM to better describe a specific case study.

Some regulatory frameworks and guidelines, such as the UWWTD revision (Woods-Ballard & Cherrier 2019) and VSA's integrated guidelines (VSA 2025), indicate a growing focus on monitoring and reporting policies, i.e. using monitoring data for compliance assessment. Incorporating a ‘reporting layer’ in the model could allow for the testing of strategies aimed at maintaining data quality and regulatory compliance. Finally, considering the influence of public interest, as seen in recent developments such as the UK water scandal (Usher 2023) could help the model reflect societal responses to water quality issues in the context of CSOs.

CSOs contribute substantially to wet weather pollution and current sensor technology enables detailed monitoring to support system optimization, compliance assessments, and future infrastructure planning. In light of emerging policies, such as the UWWTD which mandates CSO monitoring, it is uncertain which policy instruments effectively motivate sensor adoption by wastewater operators. This study addresses this gap by developing an ABM prototype to simulate sensor adoption in Swiss CSO tanks, as a first step towards improved monitoring practices. It draws on expert elicitation and psychological theories (TPB) for behaviour and opinion exchange (BCOD) and includes stakeholders in urban wastewater management, such as WWTP operators, sub-state authorities, and municipalities.

The simulation results suggest that a regulation is the most effective policy, with an increase of around 20% of sensors installed among municipalities, while the scenarios of decreased communication and better technology yield outcomes with a difference to the baseline smaller than 10%. These quantitative results need to be taken with care since the SA highlights the uncertainty of the prototype towards initial agent parameterisation. Furthermore, the numerical experiments suggest that the uncertainties inherent to the opinion dynamics algorithm require interpretation of the results.

The prototype faces challenges regarding validation due to data scarcity. Further research should address the ABM's empirical foundation. We consider TPB suitable for modelling sensor adoption in CSOs and the collection of tailored survey data would refine agent behaviour. Collecting datasets of sensor adoption data, e.g. from CSO spill monitoring, will enable calibration and validation over time.

Despite the apparent limitations of this prototype, the development of an ABM drives data collection, and integration of different perspectives, knowledge sources, and theories. ABM can be a useful approach in cases where the complexity of the system prevents the use of traditional modelling techniques or longitudinal real-world studies. The implementation of an ABM requires concisely formulated assumptions which can then be validated or refined by domain experts. This process greatly enhances system understanding, especially for interdisciplinary stakeholder groups, such as ours, and counter-intuitive model behaviour can yield valuable insights to identify knowledge gaps.

We thank Philipp Ladner (ETHZ) for the first implementation of an ABM in Python, Reto Steinemann (Chestonag AG) for SCADA and data support, and Hans Balmer (Canton of Zurich), Reto Battaglia (Canton of Berne and VSA), and Prof. Max Maurer and Kukka Ilmanen from ETHZ/Eawag for their valuable feedback.

This work was supported by BAFU through the POLAAR project (Grant No. 16.0070.PJ/R182-1359).

S.D. performed the literature review, designed the detailed research framework, refined the research questions, employed the methodology, constructed the systems model, collaborated with key stakeholders, analysed the simulation results, and wrote the original draft and the final version of the manuscript. He did most of the work on the documentation. L.M. designed the conceptual framework and contributed to the detailed framing of the study and contributed to the draft and final version of the manuscript. J.P.C. provided advice on system dynamics modelling and details of the ABM framework and contributed to the draft and final version of the manuscript. He helped with the implementation and provided detailed guidance for the documentation. M.F. provided input and ongoing advice at all phases of the project regarding Swiss water governance, regulation, and policy assessment. Contributed to the draft and final version of the manuscript. J.R. drafted the initial conceptual framework, conducted the initial background research, engaged key stakeholders, provided the tools and resources, secured the funding, and wrote the draft and final version of the manuscript. All authors have read and agreed to the published version of the manuscript.

Participating experts in this study provided non-sensitive data. While no ethics committee approval was required due to the nature of the data, informed consent was obtained from all participants involved. No sensitive data, such as survey results, is published on Zenodo.

All relevant data are available from an online repository at https://doi.org/10.5281/zenodo.4817329.

The authors declare there is no conflict.