DEVELOPMENT OF A DIGITAL TWIN-BASED DECISION SUPPORT PLATFORM FOR DRINKING WATER QUALITY MONITORING AND MANAGEMENT

CCWI 2025 - 21st Computing & Control for the Water Industry Conference, Sheffield, UK DEVELOPMENT OF A DIGITAL TWIN-BASED DECISION SUPPORT PLATFORM FOR DRINKING WATER QUALITY MONITORING AND MANAGEMENT Akrivi Alexandraki1 , Grigorios Kyritsakas1 , Alex van der Helm2 , Panagiotis Kossieris3, Christos Makropoulos3, Luuk Rietveld1 1 Department of Sanitary Engineering, Delft University of Technology, Stevinweg 1, Delft, 2628 CN, the Netherlands 2Waternet, Korte Ouderkerkerdijk 7, 1096 AC, Amsterdam, the Netherlands 3 Department of Water Resources and Environmental Engineering, School of Civil Engineering, National Technical University of Athens, Heroon Polytechneiou 5, 15780 Zographou, Greece 1Corresponding author: A.A.Alexa[email protected] ABSTRACT Climate change, ageing infrastructure, and the revised EU Drinking Water Directive highlight the need for proactive, risk-based approaches to drinking water management. This study presents the development and deployment of a modular, FIWARE-compliant decision-support platform aimed at enabling realtime risk assessment and operational decision-making in drinking water treatment. Implemented at Waternet’s Leiduin Drinking Water Treatment Plant, the platform integrates real-time data from SCADA and machine-learning-based soft sensors to monitor key treatment processes. It quantifies the likelihood and severity of short-term hazards—such as high turbidity and insufficient ozone exposure during ozonation (low contact time, CT)—as well as long-term risks, including saturation of granular activated carbon (GAC), and elevated assimilable organic carbon (AOC) levels, which can promote microbial growth in the non-chlorinated distribution network. By generating timely alerts and supporting risk-informed decisions, the platform helps operators and water technologists take preventive actions to maintain providing high quality drinking water. Keywords: Decision-making platform, Water Quality Monitoring, Risk Assessment INTRODUCTION Drinking water companies are increasingly challenged by climate change, ageing infrastructure, and evolving regulatory frameworks such as the revised EU Drinking Water Directive (DWD) [1]. Climate-related hazards—including droughts, extreme rainfall, and temperature shifts—can degrade source water quality by increasing turbidity, microbial loads, and chemical contamination, while also placing stress on hydraulic infrastructure and supply reliability [2]. In parallel, ageing infrastructure is surpassing its design lifespans, leading to higher maintenance needs, reduced treatment efficiency, and increased risks to public health and service continuity [3]. The revised DWD mandates a transition toward preventive, risk-based management approaches and requires improved transparency, monitoring, and data integration across the supply chain [1]. To address these challenges, the water industry is transitioning from reactive monitoring to real-time, data-driven frameworks supported by advanced sensing CCWI 2025 - 21st Computing & Control for the Water Industry Conference, Sheffield, UK capabilities and hydro-informatics analytics. Modern systems—such as datadriven soft sensors—leverage data from SCADA systems, IoT devices, and laboratory analyses, along with advanced machine learning techniques, to generate predictive insights into key treatment processes [4], enabling proactive and informed decision-making. To facilitate the transition toward anticipatory and risk-based drinking water management, this study presents the development of a modular decisionsupport platform, designed within the framework of the ToDrinQ Horizon EU project [5]. The platform is currently being developed as a demonstration case at the Leiduin drinking water supply system, operated by Waternet - the only public water company serving Amsterdam and its surroundings, responsible for the entire water cycle, including drinking water production, wastewater treatment, and flood protection. The platform has been engineered to enable real-time process monitoring, probabilistic risk assessment, and data-informed operational decision-making. It integrates soft sensors that predict critical treatment parameters—including turbidity fluctuations, disinfection performance based on ozone exposure (CT) (ozone concentration × ozonation tank residence time), granular activated carbon (GAC) saturation percentage and assimilable organic carbon (AOC) levels. The platform is FIWARE-compliant, meaning it adheres to the FIWARE architecture—a standardized, open-source framework designed to support the development of smart applications through the use of interoperable data models, context brokers, and application programming interfaces (APIs) [6]. Specifically, the system leverages FIWARE components to achieve semantic interoperability between heterogeneous data sources, enable scalable data exchange via the next generation service interface (NGSI), and ensure compatibility with external platforms and legacy infrastructures [6]. NGSI is a standard API specification used within FIWARE to manage and exchange context information in a structured, real-time manner. It allows different system components to publish, query, and subscribe to dynamic data streams in a consistent format, thereby supporting context-aware decision-making. This architectural alignment facilitates seamless integration into existing operational environments while supporting the broader goals of the revised EU Drinking Water Directive. The novelty of the platform lies in its FIWARE-compliant, modular architecture, which enables secure and scalable real-time integration of soft sensor predictions into operational workflows. Its unified API design allows easy deployment within legacy infrastructures and supports future expansion with both digital and physical sensor data. METHODS Platform Integration and Architecture The decision-support platform has been designed to operate within Waternet’s secure infrastructure, ensuring interoperability, modularity, and compliance with FIWARE principles. The core of the system is a Dockerized FastAPI [7] application that packages the machine-learning-based soft sensor models— each predicting key water treatment or distribution indicator water quality CCWI 2025 - 21st Computing & Control for the Water Industry Conference, Sheffield, UK parameters—within a single unified API. This design addresses Waternet’s request for a single point of integration that is scalable and easy to maintain. For now, the platform integrates five soft sensors (Figure 1), each developed to support proactive decision-making by predicting critical operational parameters at key stages of the drinking water supply system. These include: (1) a turbidity soft sensor prediction at the intake, being the inlet of the pre-treatment stage, which informs the operators for high turbidity in the inlet so as to increase coagulant or stop intake if it is possible or needed; (2) a turbidity prediction soft sensor at the outlet of the coagulation–flocculation process, designed to complement existing turbidity measurements across all pre-treatment lanes. While turbidity is already monitored, the soft sensor adds value by enabling early detection of sensor malfunction and by serving as a training signal for coagulant dose models; (3) CT soft sensor that estimates disinfection capacity based on ozone dosage and contact time. This tool supports process optimization by helping operators balance effective disinfection with minimal ozone use; (4) a GAC saturation soft sensor, which estimates the saturation level of carbon filters, guiding regeneration planning and minimizing breakthrough risk; and (5) an AOC soft sensor, which estimates AOC levels at the outlet of the drinking water treatment plant, supporting microbial risk management in the non-chlorinated distribution network. All soft sensors are encapsulated in a unified FastAPI application, with each model containerized using Docker to ensure modularity, maintainability, and secure deployment within Waternet’s operational infrastructure. Input data are received via standardized REST API calls [8]—stateless HTTPS-based requests following the representational state transfer (REST) architectural style, which enables structured communication between distributed systems through common web operations such as GET and POST. This allows Waternet’s internal systems to integrate seamlessly and in real-time basis with the platform, obtain model predictions, and integrate the results into operational workflows without tight system coupling. The architecture supports interoperability and is extensible for NGSI-LD–based smart water networks and future integration of hard sensors. The overall data flow and system connectivity are illustrated in Figure 2. Each soft sensor model also feeds into a centralized Risk Assessment Module, which interprets the outputs to support short-term alarms and long-term planning. Details on the risk evaluation framework are provided in the dedicated section below. Deployment and Data Flow The platform only incorporates soft sensor models deployed in a Dockerized FastAPI API. Waternet’s Process Information Management System (PIMS) collects high-frequency SCADA/process data and transmits it to the API via secure HTTPS POST requests. The API receives this data, runs the soft sensor models, evaluates operational risks, and returns the predictions and alerts to PIMS. This one-way data flow ensures that the API remains passive, with no direct connection to PIMS, maintaining full internal control and security within Waternet’s infrastructure. CCWI 2025 - 21st Computing & Control for the Water Industry Conference, Sheffield, UK FIWARE Compliance and Extensibility Although the platform does not yet use NGSI-LD directly, its architecture is compliant with FIWARE design principles: open APIs, modular microservices, and containerized deployment. Future integration with components such as Orion-LD and NESSIE [9], a modular, GIS-enabled digital platform designed to acquire, process, store and manage data, is planned, enabling the platform to evolve into a fully FIWARE-native ecosystem. This ensures long-term compatibility and scalability across other utilities and contexts. Risk Assessment Module The Risk Assessment Module operates on the outputs of the soft sensor models fed with process data from Waternet’s PIMS to evaluate operational risks and support decision-making through a probabilistic approach, where risk is defined as the product of likelihood and severity. Likelihood is rated on a scale from 1 (very unlikely) to 5 (very likely), based on input from Waternet, while severity is pre-assigned a value between 1 and 5 depending on the type of hazard, to reflect the impact of an event on the water delivered to consumers. The module distinguishes between short-term and long-term risks: short-term risks require immediate operator intervention, such as when turbidity exceeds critical thresholds or CT values fall below disinfection standards, triggering alarms and proposing predefined mitigation strategies (e.g., increasing coagulant dosage); long-term risks reflect evolving trends, such as gradual saturation of GAC or increasing AOC levels, and are communicated as notifications for planning and optimization. Operators and water technologists can interact with the platform by selecting specific (soft) sensors for monitoring, viewing predictions alongside risk levels, and accessing a Summary Table of Risks (Table 1), which ranks short-term risks from highest to lowest priority using a severity scale from 5 (most critical) to 1 (least critical). Table 1 illustrates how risks are prioritized and compared, helping operators respond first to the most severe threats while maintaining a consistent baseline of high drinking water quality. Figure 1. API Structure & Data flow: Soft sensor predictions feed into the risk assessment module, which includes shortand long-term predictions. CCWI 2025 - 21st Computing & Control for the Water Industry Conference, Sheffield, UK Table 1. Summary table that operators and water technologists can see with the risks - from most urgent to less urgent – and the corresponding mitigation strategy/ies. Risk ID Sensor Likelihood Severity Risk Mitigation Strategy R1 Turbidity (Coagulation – Flocculation) High: 5 3 5 × 3 = 15 Critical Risk Adjust coagulant dose or stop intake. R2 CT (Ozonation) High: 2 5 2 × 5 = 10 Moderat e Risk Adjust ozone dose. Figure 2. The forward and backward data flow between PIMS, the connector layer, and the Dockerized API. CONCLUSIONS This study presented a modular, FIWARE-compliant platform designed for realtime, risk-based drinking water quality management that is in development for testing at Waternet’s Leiduin drinking water supply system. While initial results are based on virtual data, the platform is ready for live data integration to support operator and water technologists’ decision-making. Ongoing work includes calibration with real data, use interface enhancements, and NGSI-LD integration. CCWI 2025 - 21st Computing & Control for the Water Industry Conference, Sheffield, UK Figure 3. GAC Soft sensor using synthetic data. ACKNOWLEDGMENT Some language and grammar refinements in this manuscript were supported using AI tools (OpenAI’s ChatGPT). The authors remain solely responsible for the content, conclusions, and integrity of the work. REFERENCES [1] European Commission, “Directive (EU) 2020/2184 on the quality of water intended for human consumption (recast),” Official Journal of the European Union, Dec. 2020. [Online]. Available: https://eur-lex.europa.eu/legalcontent/EN/TXT/?uri=CELEX:32020L2184 [2] A. T. Asresu, E. Furlan, F. Horneman, F. Zennaro, N. D. Nguyen, S. Torresan, A. Critto, and A. Marcomini, “A systematic review of climate change impacts on water quality in transitional environments from a multi-hazard perspective,” Estuarine, Coastal and Shelf Science, vol. 317, p. 109194, 2025. [Online]. Available: https://doi.org/10.1016/j.ecss.2025.109194 [3] R. B. Sowby, S. Hall, T. Peterson, M. Dwiggins, C. Ball, and P. Oldham, “A policy toolbox for aging water infrastructure,” Water, vol. 17, no. 9, Art. no. 1277, 2025. [Online]. Available: https://www.mdpi.com/2073-4441/17/9/1277 [4] G. Kyritsakas, Exploring Machine Learning Applications for Improving Drinking Water Quality, Eng.D. thesis, Dept. of Civil and Structural Engineering, Univ. of Sheffield, Sheffield, U.K., Oct. 2021. [5] ToDrinQ Project, “Toolkit for adaptable, resilient installations securing high quality drinking water,” 2024. [Online]. Available: https://todrinq.eu/ [6] FIWARE Foundation, “FIWARE Catalogue,” 2022. [Online]. Available: https://www.fiware.org/developers/catalogue/ [7] S. Ramírez, “FastAPI in Containers - Docker,” FastAPI Documentation. [Online]. Available: https://fastapi.tiangolo.com/deployment/docker/. [Accessed: Jun. 20, 2025]. [8] C. Pautasso, O. Zimmermann, and F. Leymann, “RESTful Web Services vs. ‘Big’ Web Services: Making the right architectural decision,” in Proc. 17th Int. Conf. World Wide Web (WWW), Beijing, China, Apr. 2008, pp. 805–814. doi: 10.1145/1367497.1367606. [9] N. Amorsi, S. Deveughele, L. Echeverria, R. Elelman, P. Kossieris, F. Lopez, C. Makropoulos, C. Nanu, J. Pocock, V. Polychniatou, S. Siauve, A. Tejado, L. Vamvakeridou-Lyroudia, and A. van der Helm. Fiware4water: Technological and non-technological dimensions of digital water. Deliverable No.6.8, Horizon 2020 Research and Innovation Programme under Grant Agreement No.821036, May, 2022. Accessed: Feb. 14, 2025.