Digital Twin Applications in the Water Sector: A Review

Abstract

As cities develop and resource demands rise, the water sector faces crucial challenges to deliver reliable, sustainable, and efficient services. Digital Twins (DTs), virtual replicas of physical systems, offer a promising tool to transform how we manage water infrastructure. Originally developed in the aerospace industry, DTs are now gaining traction in the water sector, enabling real-time monitoring, simulation, and predictive control of water and wastewater treatment, collection and distribution networks, and water reclamation and reuse systems. While still emerging in the water sector, DTs have shown potential to enhance operational efficiency, reduce environmental impacts, and support smarter, more resilient water management. This review study provides a comprehensive overview of current DT applications in the water sector, highlighting successful case studies, technical challenges, and knowledge gaps. It also explores how DTs can help bridge the water–energy nexus by optimizing resources utilized across interconnected systems. By synthesizing recent advances and identifying future research directions, this paper illustrates how DTs can play a central role in building sustainable, adaptive, and digitally-enabled water infrastructure.

1. Introduction

Increasing population and resource consumption, urbanization, industrialization, and the impacts of climate change are putting extra pressure on municipalities to move toward more sustainable and efficient infrastructure [1]. Although wastewater treatment and water reclamation provide sustainable solutions to sanitation and water scarcity, they are facing dual challenges in this regard. Providing clean water for the continuously growing population and planning under heightened uncertainty in raw-water availability on the one hand, and the energy- and carbon-intensity of water treatment, water distribution, and wastewater handling on the other hand, pose complex challenges for management and planning in the water sector [2]. In this context, employing advanced tools that combine real-time data collection, advanced data analytics, and model-based forecasts is becoming increasingly crucial for informed decision-making in the water industry. A widely discussed approach is the development of a virtual or digital representation of an actual operating system, known as a Digital Twin (DT), which satisfies the need for timely monitoring, reliable prediction, and optimized operation of water systems. This method enables more accurate simulations and predictions, leading to improved system performance and sustainability [3]. However, ongoing questions remain regarding the full integration of DTs into existing infrastructure and the challenges of scalability across water systems.

A DT is defined as the digital or virtual representation of an operating physical system [4], and has been used by the aerospace industry since the 1960s [5]. Nowadays, the concept of a DT is increasingly being adopted across a variety of industrial sectors, including manufacturing and production [6,7,8,9], smart energy systems [10], urban environment [11], residential management [12,13,14], and transportation [15]. In the past few years, DTs have been increasingly studied in the water sector, mainly for the design, control, and optimization of processes (e.g., collection, treatment, and distribution systems). While the implementation and application of DTs in the water sector is still a relatively new concept for researchers and municipalities, an increasing number of studies assessing different challenges associated with applying DTs in the water industry have been published in recent years. There are also a few recent studies on applying DTs in real-world water systems around the globe, in which a small number of cases have come very close to the implementation and operation of a full-scale DT.

This literature review aims to assess the key benefits, challenges, opportunities, and currently unanswered questions regarding the utilization of DTs to manage and optimize water systems. It explores how DTs are improving efficiency and sustainability in different water systems, identifies current limitations and gaps in their use, and discusses the future developments needed to enhance their integration and scalability. Following the introduction, essential DT terminology and modelling paradigms are established to provide a shared conceptual framework. The enabling system architecture is then described—covering data pipelines, sensor–actuator alignment, and between physical assets and their virtual counterparts. The analytical core synthesizes 147 peer-reviewed studies spanning the urban water cycle (review/conceptual papers, wastewater treatment, wastewater collection, water distribution, water treatment, water reclamation, and desalination). Moreover, emerging DT deployments in the real world are examined to demonstrate operational feasibility. Finally, this paper identifies research gaps and policy actions required to advance DT adoption across interconnected water- and energy-infrastructure systems, thereby situating the review’s findings within a forward-looking agenda for the field.

2. Search Methodology

2.1. Scope and Search Strategy

This search supports only the quantitative synthesis reported in Section 6 and Appendix C. Other sections may cite additional peer-reviewed sources outside the 2015–2025 window for context. All counts and trend analyses in Section 6 are derived from a bounded corpus of 147 peer-reviewed studies.

The Web of Science Core Collection and Google Scholar were interrogated to maximize coverage of peer-reviewed engineering and environmental journals. Due to digital twin still being an emergent terminology in the water sector, the exact phrase “Digital Twin” was combined with the following sector-specific terminologies: “wastewater treatment”, “sewer network”, “wastewater management”, “sewer systems”, “sewer management”, “sewer pipes”, “wastewater system”, “water distribution”, “water treatment”, “water resource recovery”, “water reuse”, and “desalination”; yielding 12 reproducible queries.

The search was restricted to publications dated 1 January 2015–1 May 2025. Records were retained if they (i) were peer-reviewed journal articles or full conference papers written in English; (ii) described, validated, or critically analyzed a DT or a DT-enabled framework applied to a real-world water infrastructure (i.e., treatment, distribution, collection, reuse, or desalination); and (iii) reported original data, models, or substantive insights. Conceptual commentaries, editorials, general modelling studies that did not include any identifiable DT element (e.g., a virtual model continuously updated with sensor data, a cyber-physical synchronization loop, or a DT-based decision-support framework), and papers outside the water domain were excluded. Review and conceptual papers inform the landscape synthesis but are not counted as implementations. The final corpus comprised 147 studies; however, domain categorizations in Section 6 allow a study to appear in more than one topical bucket when warranted; therefore, domain totals can exceed N = 147.

2.2. Extraction and Coding

For each included study, we extracted bibliographic metadata and structured fields: article summary, country/region, application segment, DT components present (per Appendix A), publication year, and related study-level tags. These records are available in (Supplementary Material (Table S1_Paper_Information.xlsx).

DT components were coded as present/absent (1/0) based on explicit evidence in the paper, following the definitions in Appendix A. We also applied a binary machine learning (ML) tag. A study was labeled ML = 1 if it implemented ML methods for calibration, prediction, classification, soft sensing/state estimation, control, or surrogate modeling (e.g., regularized regression, tree ensembles, gaussian processes, neural networks, reinforcement learning).

3. Definitions and Fundamentals

3.1. Model and Simulator

Modeling and simulation are fundamental components of any decision-making or management system that requires a virtual entity, such as a DT. Modeling physical, chemical, and biological processes helps acquire insights into these processes, develop control strategies, predict behavior of the system under various operational conditions, and optimize the desired outcomes [16,17]. In the context of the water industry, three primary modeling approaches are used, including mechanistic, empirical, and hybrid models.

Mechanistic modeling approaches are well-established across many industries [18]. The approach has been widely applied to wastewater treatment and water resource recovery schemes, as some important elements of urban water management, in the past decades (starting with Andrews, 1968 [19]). Mechanistic models are a representative description of a process or a set of processes (i.e., a system) that exhibit a relatively small number of state variables whose values support interpretation and comparison by subject matter experts. They are typically based on conservation laws (e.g., mass balance) combined with a community consensus to describe empirically observed relationships (e.g., alpha factor and Monod equation). Mechanistic models are typically expressed as a set of mathematical equations [20]. An example of a mechanistic model is the settling velocity function and settler model developed by Takács et al. in 1991 [21]. This model satisfies mass conservation while providing an empirical description of the settling dynamics.

In contrast, empirical models are focused on the description of empirically observed relationships and do not explicitly satisfy conservation laws. Empirical models can handle noise in measurements and identify patterns without prior knowledge of underlying processes, making them useful for predicting more complex systems such as achievable water quality in a multi-stage treatment system [22]. They are also faster and more cost-effective to develop, whereas mechanistic models require computational resources and expertise [23]. However, the structure of empirical models may be influenced by basic knowledge on the data-generating system. For example, the selection and preprocessing of measurement time series may be based on a prior qualitative understanding of the modeled process. Similarly, the physical location of a sensor (e.g., upstream or downstream of a treatment process) or basic understanding of the direction of causality may influence the allocation of the corresponding signal as a model input and output. Empirical models are used to describe, predict, and forecast patterns observed in the data.

The concept of combining mechanistic and statistical models started in the early 1990s, and it was initially used as a hybrid model in the field of chemical engineering [20]. The concept mainly emphasizes the use of both mechanistic and empirical models, highlighting the incorporation of all available knowledge (including information, data, and domain expert knowledge) into the developed model to strengthen the outputs by taking advantage of both types of models.

Computational and dynamic models developed based on mathematical equations (e.g., algebraic or partial differential equations) to describe physical or chemical phenomena involved in the process, can be integrated into simulators that enable forecasting the behavior of a system. The majority of simulators developed in the water industry are used primarily for planning infrastructure development, system optimization, and resource allocation and require significant modifications to qualify for use in real-time operational contexts [24]. For an operational DT, the underlying model should (i) represent the decision-relevant states and outputs that drive actions (e.g., effluent quality, tank levels/pressures, energy use) and map them to available sensors; (ii) include the actuation pathway and constraints (e.g., blower speed, pump variable frequency drives, valve status; physical/safety limits); (iii) represent key disturbances and boundary conditions (e.g., influent flow and load, temperature, demand), with short-range forecasts where applicable; (iv) support state/parameter synchronization via data assimilation or adaptive calibration so that the virtual state tracks the plant; and (v) quantify uncertainty and latency requirements so the model can deliver predictions/decisions within the update interval required by the use case, with validation/verification commensurate with that purpose (see Appendix A).

Currently, most simulators cannot accurately capture real-time process states at a high level of fidelity, often required for implementation of a DT [25]. Examples of simulators that are commonly used for process simulation in the water industry include EPANET, SUMO, and BioWin. For DTs, these simulators must be wrapped with (a) live data adapters, (b) estimation/assimilation to keep states current, (c) explicit actuator/constraint mappings for safe decision making, and (d) a runtime scheduler that meets operational latencies. Without these layers, such tools remain planning simulators or, at most, enable a digital shadow (DS) rather than a full DT.

3.2. Digital Twin and Digital Shadow

A DT system is required to have a physical entity, high-fidelity simulator, physical sensors and analyzers for the physical entity, actuators for virtual connection, physical-to-virtual connection, advanced data analysis, and interaction and services interface to serve its full practical applications. Detailed descriptions for each of these components are provided in Appendix A.

When deployed at full scale, DTs offer advantages over traditional design and decision-making methods. Unlike conventional offline models, which are typically static and rely on fixed assumptions or historical datasets, DTs remain continuously synchronized with the operational system through real-time data streams. This dynamic linkage means that design or upgrade scenarios can be evaluated not only in theory but also under conditions that reflect the system’s current and evolving state [7]. By representing an operational system, DTs facilitate rapid prototyping and what-if analyses for performance prediction and risk assessment [26]. Because they maintain a high-fidelity, dynamic, self-evolving model, DTs support improved maintenance, anomaly detection, performance tuning, sensitivity analysis, auto-calibration, and overall system optimization. They also enable training and informed decision-making under unusual operating conditions, as well as remote or automated control. Altogether, these capabilities drive more sustainable and efficient operations [5].

Figure 1 shows the typical structure of a DT in the water industry, with the major components that are required to fulfill its purposes.

As mentioned previously, a simulator model is often used for characterization and planning purposes, mainly to minimize the costs and labor before implementation of the full-scale plan (e.g., expansion and process modifications). When integrated into advanced digital frameworks, these models can evolve into more sophisticated systems such as DSs or DTs. The key difference between DS and DT lies in data flow directionality [27]. In a DS, data from the physical asset automatically updates the digital model, but any feedback from digital to physical remains manual or absent. In contrast, a DT supports automatic, bidirectional data: real-time data from the physical object updates the virtual model, and the virtual model can autonomously trigger in the physical system [28]. Importantly, “real-time” should not be treated as a fixed latency; the acceptable update interval depends on the twin’s purpose (e.g., fault detection vs. closed-loop control) and on the governing process time scales (e.g., aeration dynamics vs. sludge-age) [29]. Thus, a DT is defined by automated data ingestion and dynamic model updating, while the specific update cadence is application-dependent rather than universal.

In a DS, the virtual model’s outputs require manual processing, limiting its use in strategic planning [5]. A DT, however, consists of an operational physical entity, a virtual entity that mirrors its current state and behavior, and a closed-loop, two-way data exchange (see Figure 2). The virtual entity continuously synchronizes with real-world data, enabling real-time feedback, predictive insights, and remote or automated control under varying operational conditions.

4. System Integration

As discussed above, two-way communication is an essential component of the conception of DTs in the water sector. Moreover, the need for trust and reliability of automated decisions motivates the use of mechanistic models as the basis of DTs, even if augmented with ML tools. Thus, it is critical that the states in the model are correctly paired with the corresponding actuator and sensor signals in the physical system. Additionally, since most industrial-scale systems in the water sector are one-of-a-kind, the mapping between the physical system components and the digital entity is typically established on a case-by-case basis. This tends to be a manual and tedious process for the reasons discussed below. In the present work, frequent practical issues are addressed and recent efforts that may facilitate this task are highlighted.

4.1. Variable Naming in Process Flow Diagram (PFD), Process and Instrumentation Diagram (P&ID), Data Acquisition and Control (DAC), and Model

In most pilot- and full-scale modeling projects, information that is useful for the conceptualization of a DT resides in multiple locations. Information is typically collected from a variety of isolated sources, such as flow sheets or process flow diagrams (PFDs), process and instrumentation diagrams (P&IDs), historians (time-series databases), laboratory records, and stand-alone process models, each with its own notation and level of detail. Flow diagrams provide a high-level overview of major processes and some measured variables; however, they lack standardized conventions and often require manual interpretation. P&IDs (e.g., using ISA 5.1 symbols [30]) identify actuators, sensors, and control loops with compact icons and single-letter variable codes (e.g., “F” for flow rate), yet industry standards do not consistently cover water quality variables (e.g., electrical conductivity and pH), leading to ambiguity. Historians and supervisory control and data acquisition (SCADA) systems store sensor and actuator signals as “tags” or registers and typically offer little contextual information about their physical location or signal type, making it difficult to match the tags to the processes they monitor. Laboratory data, collected, prepared, and analyzed through manual or semi-automated procedures, are essential for DT accuracy but rarely come in a standardized format listing sampled variables and locations. Finally, existing process models (not yet integrated into a DT) use proprietary shorthand to reference variables, often differing from P&ID and SCADA notations. Although efforts have been made to standardize wastewater treatment model notations [31], thus far, no single authority enforces consistency across all water sector diagrams and databases.

The current work’s findings suggest that consulting all sources of information listed above is required to establish a complete list of variables, measured or not, and to identify the correspondences between the data acquisition and control system (i.e., programmable logic controller (PLC) or SCADA) and the variables represented in the model. Today, this is a manual, tedious, and possibly error-prone task. A more efficient and robust approach to identifying the correspondence is needed, especially for small-scale systems. To our knowledge, two main approaches have emerged to streamline this process. First, a simple “ground truth” method uses the P&ID as the authoritative reference: variable names in the P&ID are manually matched with SCADA tags, laboratory spreadsheets, and model notations. Once this mapping is agreed upon, all other representatives, including DT, must adhere to the ground truth. Although robust, this approach still relies on human effort for maintenance and updates. Second, computer-aided knowledge management leverages a centralized knowledge graph or ontology to consolidate information from flow diagrams, P&IDs, SCADA, and models into a single framework. Such a knowledge graph can automate validation of the DT’s structure and function, an approach already demonstrated for automated model generation in manufacturing [32], and promises greater transferability of DT platforms between water sector plants, though its cost–benefit profile has not yet been fully evaluated.

4.2. Data Pipelines

In addition to the crucial matching of input and output variables in the model with actuators and sensors on the physical plant, additional meta-data will be necessary to ensure the proper design and functionality of the DT. Based on the meta-data types and recommendations listed by the International Water Association (IWA) Task Group on Meta-data Collection and Organization [33], the following is considered essential in the construction of a DT.

Every actuator and sensor signal must include its unit of measurement, converted as needed to Système International (SI) units, and a timestamp synchronized to the model’s simulated time, ideally in Coordinated Universal Time (UTC). DT’s ability to capture system dynamics also depends on the numeric and temporal resolution of data, the physical placement of sensors and actuators, and the mode of signal transmission (i.e., analog, digital, or manual records). Any preprocessing (such as aggregation, filtering, compression, imputation, interpolation, time synchronization, data fusion, or use of soft sensors) can further influence accuracy. Finally, metadata should document known sources of measurement error, including calibration drift, electrical glitches (outliers), and sensor fouling or scaling. Collecting this meta-data is essential to assess and ensure the DT’s accuracy and fitness for its purpose. These meta-data recommendations align with practices observed in real-world deployments: full-scale DT/DS implementations integrate SCADA tags and laboratory records with hydraulic or process models (e.g., Valencia water distribution network (WDN) [3], Lushan water supply [34]), underscoring the need for consistent units, timestamps, and documented preprocessing.

4.3. Data Access Rights

A successful deployment of a DT, including a two-way information flow between the DT and the physical entity it describes, will require that the plant is equipped for real-time communication and that read and write access rights are established and implemented correctly. While this offers no technological challenge, it is often a plant-internal governance bottleneck. Here are the typical questions that must be addressed:

1.

Communication infrastructure: Is the physical infrastructure ready to enable communication to, and from, the DT? For example, can the laboratory data be accessed automatically?

2.

Communication configuration: What are the requirements for granting read access to the information contained in the SCADA system, present as laboratory data, or produced by the DT?

3.

Write-access rights: Can the DT manipulate variables (e.g., setpoints) in the physical entity? If so, which ones, at what frequency, and within what range?

Consistent with these governance considerations, several full-scale systems in our inventory currently operate in advisory-only mode (e.g., Eindhoven water resource recovery facility (WRRF) [35], Gothenburg collection [36], Águas do Porto [37], Valencia WDN [3]), even where the communication stack exists.

4.4. Constraints for Optimization Based on a DT

A DT provides a flexible tool to test manipulations of physical entity in a virtual environment. However, to ensure the proposed control actions match what is feasible and permitted on the physical plant, it is important that the DT is equipped with all known constraints. Such constraints fall into two broad categories that are discussed here.

First, physical constraints define equipment limits. For instance, a valve’s opening setpoint must lie between fully closed and fully open. Second, safety or regulatory constraints restrict operations to protect the environment, equipment, and personnel, even if equipment could physically exceed these bounds. Many SCADA systems and PLCs impose such constraints automatically so that safety is guaranteed, even if the DT does not account for them. For example, a valve, while physically enabling a complete closing, may only be allowed to close down to 10%. To ensure that (a) the DT accurately reflects the behavior of the plant (monitoring and live updating) and (b) that the optimized control actions provided by the DT can be realized practically, it is important that they are accounted for during optimization.

A practical challenge is that constraint logic often resides within proprietary PLC or SCADA programming, such as variable-frequency drive controllers or third-party custom code, that utility staff may not have direct access to. Because these hidden control loops govern permissible setpoints and safety limits, failing to represent them in the DT can lead to proposed actions that cannot be executed or that violate safety requirements. Despite this, few systematic methods currently exist to extract and encode both physical and safety constraints from the plant’s control infrastructure into the DT, making it difficult to guarantee that model-based recommendations can be implemented in practice.

5. Digital Twin Applications

Several review studies have explored DT development and requirements across industries. For instance, Lim et al. (2019) [38] reviewed DT programming, applications, and benefits, with an emphasis on the manufacturing field. Similarly, Tao et al. (2018) [7] reviewed the history of DT in the real world and in the literature. They focused on several major areas where DTs are used: product design, production, prognostics, health management (the most popular area), and service. They identified two promising areas for DTs that include dispatching optimization and operational control. And finally, Rasheed et al. (2020) [39] provided a general overview of what DTs are, and what the values, challenges, and enablers of implementing and operating a DT from a modeling perspective are. Although the study was not directly related to one specific industry, some of the findings can be applied to any corresponding field, including the water industry. Compared with broad industrial reviews, water-sector applications emphasize plant-wide process control and energy optimization at WRRFs (e.g., Changi water reclamation plant [29]; EWE Cuxhaven [40]) and district-scale hydraulics for flood risk and leakage in collection/distribution systems (e.g., Gothenburg [36]; Valencia [3]). These emphases shape model choices (hybrid/mechanistic + ML) and data orchestration across SCADA and smart-meter infrastructures.

According to the abovementioned studies, although a full-scale DT, when all the required components and elements are implemented in the design, provides all the benefits and advantages of having a virtual representation of an operational physical system, some DTs are designed to serve a specific purpose. Accordingly, a DT can fall into one specific category, including Monitoring DT (MDT) [41], Optimizing DT (ODT) [42], Forecasting DT (FDT) [43], Sensitivity DT (SDT) [44], or a Wrapper for process safety [45], which are discussed in detail in Appendix B. While these categorizations offer a focused application of DT technology, it is important to recognize that they may not fully align with the holistic definition of DTs that integrates all key functions into a unified system. In this review, MDT/ODT/FDT/SDT/Wrapper are used as functional roles (not mutually exclusive) that a deployment may emphasize; they do not redefine what a DT is. As DTs are ideally intended to be comprehensive digital counterparts of physical entities, categorization of them into specific types may limit the scope of their application and potentially restrict their full potential. For instance, in the concept of MDT, Wang et al. (2020) [41] utilized a DT for real-time monitoring of a system; however, the capability for data transmission or modification from the digital dashboard back to the physical system was not implemented. This approach, therefore, aligns more closely with the concept of a DS, as it lacks the bidirectional communication characteristics essential to a comprehensive DT.

The segmentation into singular-purpose DTs, while valuable for specific operational goals, might miss the opportunity to leverage the full range of interconnected capabilities that a DT can provide when implemented as an all-encompassing system. Therefore, although these specific types of DTs contribute significantly to the advancement and specialization of technology, it is essential to acknowledge that a more integrated approach might offer a richer and more accurate representation of what DTs are fundamentally designed to achieve. This perspective encourages further exploration into ways these different functionalities could be brought together into a more comprehensive DT model.

6. Digital Twins in the Water Industry

A review of 147 studies published between January 2015 and May 2025 highlights the growing maturity of DT applications in the water sector (see Figure 3). Research has expanded from a single publication in 2015 to 41 in 2024. These studies are grouped into seven topical categories that are most applicable to the water industry: (1) review and conceptual studies, (2) wastewater treatment, (3) wastewater collection systems, (4) water distribution networks, (5) drinking water treatment, (6) water reclamation, and (7) water desalination. Most studies focus on water distribution networks (a total of 47 studies), wastewater treatment (a total of 41 studies), and drinking water treatment plants (a total of 25 studies). Additional topics include reviews and conceptual frameworks (a total of 38 studies), wastewater collection (a total of 17 studies), water reclamation (a total of 17 studies), and desalination (a total of 4 studies). Coverage remains uneven across domains. Desalination is scarcely represented (4 studies), with water reclamation and wastewater collection also underexplored (17 each); these areas merit focused DT research and deployment, consistent with the identified gaps in membrane-based systems. Because several studies spanned more than one domain, multi-label classification was used and each study assigned to all applicable categories; consequently, category totals reflect 189 assignments across 147 unique studies. While the scope of the topic is expanding, key gaps remain, particularly in membrane-based systems and applications in low- and middle-income regions.

The coded corpus is geographically concentrated in higher-income regions. Based on the country/region tags compiled in our screening sheet (see Section 2.2 and Supplementary Material Table S1_Paper_Information.xlsx), China and the United States contribute the largest shares (n = 18 each), followed by the United Kingdom and Italy (n = 6 each), with Singapore and Canada close behind (n = 5 each). Representation from low- and middle-income regions is comparatively limited, which likely reflects differences in sensing/IT infrastructure, metadata discipline, and governance around control authority. This imbalance should be considered when generalizing results; it also motivates future work pairing lower-cost sensing and open data stacks with staged automation tailored to utilities operating under tighter resource constraints.

Most DT prototypes described in the literature implement only part of the eight components defined in Appendix A. Early work (2015–2018) primarily focused on detecting anomalies and developing soft sensors. Since 2019, studies increasingly combine live control-system data with calibrated simulations. By 2023, a few utilities had begun to implement full model-predictive control systems. When a study applied supervised/unsupervised ML, deep learning, reinforcement learning, or hybrid physics–ML approaches, it was labeled as an ML study; routine statistical fitting or purely mechanistic calibration without a learning algorithm was not labeled ML. 58 of 147 studies (40%) apply ML methods, indicating growing adoption of hybrid modeling. However, only eight studies contain all the defined components essential for a fully implemented DT, which will be discussed in the next section along with other full-scale DTs implemented in the water industries. A comprehensive review was conducted for these research studies and additional information is provided in Appendix C.

Conceptual and review studies have systematically mapped the landscape of DTs in the water sector, identifying the core functions, architectures, and challenges for practical implementation. Early surveys catalogued available open-source tools and framed DTs as enablers of carbon-aware operations [2], while broader reviews highlight transferable strategies from manufacturing and energy systems [10]. Foundational works by Grieves and Vickers [46] and Moser et al. [47] defined essential DT characteristics such as dynamic synchronization, bidirectional feedback, and modular and mechanistic models, laying the groundwork for water-sector adaptations. More recent studies have proposed service-oriented architectures (referred to as Digital Water Services) that federate asset-level twins via standardized Application Programming Interfaces (APIs) [48], and cyber-physical testbeds that integrate artificial intelligence (AI), soft sensors, and threat modeling for drinking-water systems [49].

In wastewater treatment, research has progressed from standalone soft sensors and anomaly detection to integrated, model-predictive control systems. Initial efforts applied Long Short-Term Memory (LSTM) and deep-belief networks to estimate influent quality and energy use [50,51], while hybrid frameworks coupling mechanistic activated-sludge simulators with Convolutional Neural Networks (CNNs) have achieved real-time forecasting with high interpretability [52]. Pilot-scale implementations demonstrate energy-aware control: Heo et al. [53] reduced blower energy by 17% under nitrogen-ratio constraints, and Chen et al. [54] achieved <5% mean absolute percentage error in three-hour effluent forecasts with automated dosing adjustments. Modular BIM-enabled (building information modeling) approaches for pump maintenance [55,56] and federated “digital water services” architectures [48] point toward full-plant DTs that encompass design, operation, and asset management.

Sewer network DTs tackle the twin imperatives of overflow mitigation and structural health monitoring. Early empirical diagnostics used stacked autoencoders for pump-failure detection [57] and deep-reinforcement learning for energy-efficient pumping [58]. Stormwater twins integrating storm water management model and data-assimilation filters supported flood forecasting [59,60], while VandCenter Syd’s “living” DT concept distinguished real-time operational twins from offline prototypes [61]. Advanced sensing and modeling—physics-informed neural networks for heat-transfer estimation [62] and rapid LiDAR-to-3D-model workflows [63]—enable continuous structural updates. Gothenburg’s 140 km combined sewer overflow (CSO) control system and Germany’s KaSyTwin robotics initiative showcases fully cloud-hosted, predictive-maintenance DTs in action [36,64].

Water distribution DTs emphasize leak detection and energy optimization across large urban networks. Pioneering frameworks in Valencia combined EPANET hydraulic models with geographic information system (GIS) and telemetry for scenario analysis [3,65]. Lab-scale studies achieved >97% leak-classification accuracy using CNN–VAE (Convolutional Neural Network- Variational Autoencoder) architectures [66], while mixed-integer nonlinear programming models explored turbine-based energy recovery [67]. Real-time calibration routines adjust demand patterns via pressure feedback [68], and temporal graph–convolutional controllers optimize pump speeds in EPANET–WNTR environments [69]. Roadmaps for staged DT maturation outline the path from static models to self-updating, value-tracking systems [70], and service-oriented “Digital Water Service” platforms promise seamless integration across distribution, treatment, and asset management.

Drinking water treatment DTs have evolved from anomaly-detection proofs of concept to plant-wide, self-adapting control systems. Unsupervised models for cyber-physical anomaly detection [71] and BiLSTM-based (Bidirectional Long Short-Term Memory) quality alarms [72] led to predictive-maintenance twins for pumps and valves [73] and coverage-based online model updating in treatment plants [74,75]. Utility-scale frameworks combining BIM and asset management [76] and service-oriented “Digital Water Service” architectures [48,77] laid the data-governance groundwork. Recent demonstrations coupled multi-scale process twins with edge micro-services, reinforcement-learning dose controllers, virtual reality cyber-security training, and energy-optimal pump scheduling [78,79,80,81,82]. The latest hybrids dynamically switch between first-principles and LSTM surrogates, keeping turbidity prediction errors < 0.2 NTU across variable raw-water conditions [83], signaling a shift from isolated algorithms to resilient, policy-aligned decision-support twins for safe and energy-efficient drinking-water delivery.

Water-reclamation DTs address the dual complexity of advanced treatment and reclaimed-water distribution. Conceptual studies framed twins as next-generation tools for multi-barrier purification [84], while adaptive grey-box observers [85] and hybrid mechanistic/ML models [20,29] proved the value of empirical calibration. Supporting modules emerged for influent synthesis [86] and cyber-resilience assessment [87,88]. Governance roadmaps and asset-twin federation schemes preceded plant-wide, real-time twins: a hybrid neural/Activated Sludge Model (ASM) model cutting nitrate-prediction error to ≈2% [89], a two-hour-look-ahead operational twin at Eindhoven WRRF [35], and soft-sensor pipelines delivering sub-minute latency [90]. Greenhouse gas emission-aware (GHG-aware) prediction engines [91,92] and industrial-reuse salinity twins [93] broaden the focus from process reliability to sustainability metrics, pushing WRRF DTs toward integrated, emissions-conscious decision support.

Desalination DTs concentrate on energy, fouling, and cost optimization in Reverse Osmosis (RO) systems. A model-predictive controller prototype for a Peruvian pilot plant [94] and simulation reviews of RO optimization strategies [95] set the stage. Operational twins now predict specific energy consumption at full-scale [96] and support condition-based membrane cleaning and replacement during algal blooms at the Carlsbad plant [97]. These studies illustrate how real-time soft sensors, energy/fouling trade-off analytics, and decision-support dashboards are converging into near-complete, plant-wide DTs that drive higher recovery, lower specific energy, and smarter maintenance scheduling in energy-intensive desalination facilities.

7. Digital Twin Implementation in Real-World Systems

As discussed previously, only a small subset of the 147 reviewed studies included all the components required for a true DT. In this section, we examine those studies alongside additional full- and pilot-scale implementations, drawn from global projects that have not been formally published, in detail. Although many of these efforts claim to be complete DTs, they function more like DSs due to lack of bidirectional, real-time connectivity between the physical asset and its virtual counterpart. Their digital models serve only as advisory tools and cannot autonomously adjust the actual system. From this analysis, nine full-scale (see

Table 1. Full-scale implementation of DT and DS in a Water System.

Facility Name	Location	Facility Type	Implementation	Reference
Changi Water Reclamation Plant	Singapore	Water Reclamation	DT: Operated in the advisory-only mode, but it has capability of bidirectional communication	[29]
Eindhoven Water Resource Recovery	Netherlands	Water Resource Recovery	DS: Advisory only	[35]
Bresso-Niguarda Water Resource Recovery	Italy	Water Resource Recovery	DS: Advisory only	[98]
Gothenburg Regional Sewage System	Sweden	Wastewater Collection	DS: Advisory only	[36]
EWE WASSER GmbH	Germany	Wastewater Treatment	DS: Advisory only	[40]
Águas do Porto	Portugal	Integrated Urban Water Cycle Management	DS: Advisory only	[37]
Lushan Water Supply Company	China	Water Supply System	DT: Advisory and process control	[34]
Water Distribution Network of Valencia	Spain	Water Supply System	DS: Advisory only	[3]
Strongford Wastewater Treatment Plant	United Kingdom	Wastewater Treatment	N/A *	[99]

Note: * N/A: not enough information.

) and seven pilot-scale (see

Table 2. Pilot-scale implementation of DT and DS in a Water System.

Authors	Location	System Type	Implementation	Reference
Bartos & Kerkez	USA	Surface Water Systems	DS: Advisory only	[60]
Komulainen & Johansen	Norway	Wastewater Treatment	DS: Advisory only	[100]
Lian et al.	Various communities China and Australia	Pilot plants for desalination	DS (2) * + DT (2)	[88]
Zekri et al.	Oman	Water distribution	DT: Advisory and process control	[101]
Gomez-Coronel et al.	Mexico	Water distribution	DT: Advisory and process control	[102]
Lumley et al.	Sweden (7) + Norway	Wastewater Collection	DS: Advisory only	[103]
Lumley et al.	Denmark (4) + Netherlands + Italy	Water Treatment	DS: Advisory only	[103]

Note: * Numbers in parentheses indicate the number of individual pilot plants.

) implementations emerged as genuine DT candidates. These projects are discussed in the following sections.

7.1. Full-Scale Implementation

Changi Water Reclamation Plant [29]. Jacobs partnered with Singapore Public Utilities Board to develop a DT for the Changi Water Reclamation Plant using the Replica™ platform. This DT integrates approximately 1200 data tags in real time from the SCADA system for dynamic modeling of the entire plant process. It is a hybrid model and uses both mechanistic and empirical models that are automatically calibrated through SCADA and laboratory data. It is implemented in advisory-only mode and has three primary uses: (1) evaluation of DT vs. measured results to highlight areas of attention for operational staff, (2) simulating scenarios for informed operational planning, and (3) forecasting significant process events up to 5 days in advance.

Eindhoven Water Resource Recovery Facility [35]. The Eindhoven WRRF in Netherlands utilizes a DT for dynamic plant-wide process monitoring, scenario analysis, and forecasting. The continuous data pipeline developed for the DT uses the plant’s SCADA system to collect and store raw data in an iHistorian database. Every 2 h, the data is automatically filtered and converted into various data formats for modeling. The plant-wide model aggregates the three physical parallel treatment trains into one treatment lane with the combined volume and flow rate. An important aspect of the DT involves refining the compartmentalized model of the aeration tank to better account for non-ideal mixing, which is a well-documented challenge in such systems. While DT has led to increased accuracy of predicted dissolved oxygen and air flow rate profiles, nitrate and mixed liquor suspended solids are especially susceptible to prediction inaccuracies in wet weather conditions.

Bresso-Niguarda WRRF [98]. Gruppo CAP, the integrated water service provider of the Metropolitan City of Milan, implemented the Twin Plant DT system developed by the Danish Hydraulic Institute (DHI A/S) at the Bresso-Niguarda wastewater treatment plant (WWTP) in Milan, Italy. This system automatically acquires online information from over 60 sensors, including 18 energy meters, and uses the Wastewater Treatment Plant Expert Simulation Tool (WEST) developed by DHI A/S to forecast plant performance with a 24-h horizon. Using these predictions, operators, process engineers, and energy management personnel can evaluate various operational strategies and identify optimal operational settings.

Gothenburg Regional Sewage System [36]. Gothenburg’s central treatment facility, the Rya WRRF in Sweden, experiences large variations in influent flow due to heavy rainfall and 25% of the sewers being combined. As a result, the city faces the risk of flooding and discharge of untreated wastewater during heavy rain events. The DT includes a detailed hydrodynamic model of the collection system and its 250 km² catchment. This allows for real-time prognosis, forecasting, and scenario training.

EWE Wasser GmbH [40]. German utility EWE Wasser GmbH partnered with Xylem to deploy a DT at the Cuxhaven treatment plant. The DT is connected to the plant’s SCADA system which provides real-time data as well as several virtual sensors which provide estimations of incoming constituent loading. In particular, the DT utilizes carbon, nitrogen and phosphorous loading estimations coupled with ML to predict carbon, nitrogen, and phosphorous removal in real time. These dynamic estimations of biological oxygen demand (BOD) and nutrient removal enable the tuning of aeration setpoints to minimize power consumption. Since implementation, the Cuxhaven treatment plant has been able to decrease aeration by 30% while maintaining effluent quality standards. This has resulted in an annual power saving of 1.2 million kWh.

Águas do Porto [37]. Águas do Porto, the municipal water utility of Porto in Portugal, implemented H2PORTO, a real-time integrated management platform that functions as a DT for the entire urban water cycle, covering water supply, wastewater, stormwater, and coastal systems. The system integrates data from 22 different sources, including SCADA systems, meteorological inputs, and historical records. H2PORTO uses the Bentley Systems OpenFlows Suite, including WaterGEMS for water distribution modeling, SewerGEMS for wastewater network simulation, and FLOOD for urban flood modeling. These tools provide hydraulic and hydrological modeling capabilities within DT. This DT enables real-time system monitoring, predictive analytics (e.g., for flood risks and water quality), and supports operational decision-making, emergency response, and preventive maintenance planning.

Lushan Water Supply Company [34]. Lushan Water Supply Company in China upgraded its water supply infrastructure in Guling using a DT developed in a partnership with Siemens. The system is powered by SIMATIC PCS neo, Siemens’ web-based distributed control system platform. Real-time operational data including flow rates, water pressure, and water quality metrics are collected through a network of sensors and processed via SCADA integration. The DT uses hydraulic simulation models to represent the behavior of the water distribution system. These simulations support operational forecasting, performance optimization, and asset management.

Water Distribution Network of Valencia (Spain) and its Metropolitan Area [3]. Global Omnium, the operator of Valencia’s water system, deployed GoAigua, a centralized smart water platform that acts as the DT for the city and its metropolitan water distribution network. The system collects and combines data from more than 20,000 readings, including SCADA systems, smart meters, pressure and quality sensors, and customer information systems. GoAigua incorporates advanced hydraulic modeling to simulate network behavior in real time. It provides leak detection, demand forecasting, and energy optimization tools.

Strongford Wastewater Treatment Plant [99]. Severn Trent Water, in partnership with Siemens, Atkins, Explore AI, and Xylem, is deploying Siemens’ gPROMS DT technology at the Strongford wastewater treatment plant in Stoke-on-Trent, United Kingdom. This DT integrates over 60,000 static and dynamic metrics from sensors and operational systems to model the wastewater treatment process in real time. Originally developed for the pharmaceutical and petrochemical sectors, gPROMS enables detailed monitoring of energy use, effluent quality, and GHGs—including CO₂, CH₄, and N₂O. According to this report, the project team is developing a mechanistic model of nitrous-oxide production for wastewater treatment. Funded by a £10 million Ofwat Innovation Fund award, €0.9 million from Horizon Europe, and £28 million from Severn Trent, the project aims to eliminate direct process emissions and serve as a global blueprint for carbon-neutral WRRFs.

7.2. Pilot-Scale Implementation

Pipedream: An Interactive Digital Twin Model for Natural and Urban Drainage Systems [60]. The University of Michigan developed the PIPEDREAM DT for real-time management of urban stormwater systems. This DT continuously receives data from in-situ sensors monitoring flow depth, velocity, and rainfall across the drainage network. Using a full hydraulic model based on the Saint-Venant equations, dynamically updated with a Kalman filter, PIPEDREAM simulates and forecasts network conditions several hours into the future. These forecasts enable operators and city engineers to detect flooding risks early, assess the system’s condition in unmonitored locations, and optimize control of stormwater infrastructure such as gates and detention basins.

Possible Concepts for Digital Twin Simulator for WWTP [100]. Oslo Metropolitan University and collaborators developed a DT for the Veas wastewater treatment plant in Norway. This system automatically acquires online data (e.g., flow, levels, and process variables) via SCADA, feeding a dynamic process model that simulates biological and hydraulic treatment processes. Simulations support capacity planning and performance forecasting, enabling operators and engineers to evaluate scenarios for increased inflows or regulatory changes, allowing optimization of process adjustments without requiring physical trials.

Application of Digital Twins for Remote Operation of Membrane Capacitive Deionization (mCDI) Systems [88]. A DT for several mCDI pilots in remote communities of Australia and China were developed to support maintenance planning. Real-time and historical feed and permeate data are captured via SCADA or monitoring systems. The DT employs a mathematical model of membrane aging and wear within the RO vessel, simulating degradation and cleaning cycles. Operators use this DT to autonomously control process dynamics, forecast membrane performance, schedule maintenance, and practice training scenarios.

Smart Water Management Using Intelligent Digital Twins [101]. Zekri et al. [101] proposed an intelligent DT framework for municipal water management, integrating real-time sensor data with MI-enhanced hydraulic models. For proof of concept, 100 smart water meters were installed in a residential area near Sultan Qaboos University in Oman. The DT supports system-wide optimization by enabling leak detection, operational efficiency analysis, and resource allocation. The output is used by water managers and planners to reduce losses, cut costs, and improve supply resilience under varying constraints. However, it is noted that implementation of the DT does not yet support expected functionalities. Future work is aimed at improving internet of things (IoT) failure forecasting and simulation.

Digital Twin of a Hydraulic System with Leak Diagnosis Applications [102]. The Hydroinformatics Laboratory at Instituto Tecnológico de Tuxtla Gutiérrez in Mexico developed a pilot scale distribution system coupled with a DT to diagnose leaks in water distribution networks. The DT enables remote control of the hydraulic system’s actuators via embedded microcontrollers equipped with IoT functionality. Real-time pressure head and flowrate data are transmitted to the operator interface, where a calibrated EPANET hydraulic model runs simulations to estimate pressures at nodes without sensors. To detect and quantify leaks, a genetic algorithm was developed and integrated into the system. The method was validated through various experiments, including scenarios involving multiple simultaneous leaks, demonstrating its effectiveness for real-time leak detection and system monitoring.

Connecting Digital Twins to Control Collections Systems and Water Resource Recovery Facilities: From Siloed to Integrated Urban (Waste)Water Management [103]. Future City Flow is a DT platform developed by Scandinavian utilities, universities, and DHI to optimize control of urban wastewater collection systems. It integrates a hydrological-hydraulic model with real-time data from IoT sensors, SCADA systems, and weather forecasts to continuously simulate system conditions and support proactive control of pumps and storage volumes. A web-based interface enables operators to visualize data, implement predictive control strategies, and reduce sewer overflows and flooding. Future City Flow has been validated through pilot projects in Sweden and Denmark, such as in Malmö and Lund, where it demonstrated significant reductions in overflow volumes during storm events by using forecast-driven decision support.

The second part of this study reviews TwinPlant, a DT developed by DHI in collaboration with Aarhus Vand for WRRFs. It combines a mechanistic process model (typically ASM-based) with continuous data from sensors, lab analyses, and weather inputs to support real-time monitoring, operator training, and optimization of treatment performance. Through a web-based Graphical User Interface (GUI), operators can run scenario analyses and determine optimal control setpoints for objectives such as energy efficiency or effluent quality. Pilot testing at the Marselisborg WRRF in Denmark showed that TwinPlant could effectively reduce aeration energy while maintaining stable treatment under variable conditions, validating its value for operational decision support.

7.3. Cross-Case Synthesis

Drawing on the full-scale implementations in Section 7.1 and the pilot studies in Section 7.2 (see also Table 1 and Table 2), three recurrent mechanisms explain why some projects translated analytics into process control while others remained as advisory only. First, governance and constraint mirroring consistently separated advisory dashboards from operational DTs. Where write-back authority was explicitly defined and bounded (e.g., Lushan [34]), the projected reported as “advisory + process control.” In contrast, projects that lacked transparent access to PLC/SCADA constraint logic, or where authority was not formally delegated, stayed in advisory-only DS mode despite sophisticated modeling.

Second, benefits were sustained when estimation fidelity and updating rate matched the dynamics of the target process. Pilot-scale systems typically constrained scope and instrumentation so that soft-sensor errors and data latencies were managed within a defined budget, allowing reliable tracking and, in some cases, closed-loop moves (e.g., [102]). At utility scale, exposure to wet weather hydraulics [35] and heterogeneous data streams often degraded forecast skill during transients; in these settings, teams either added hydrodynamic coupling and uncertainty reporting or remained advisory only.

Third, projects that invested early in unambiguous mappings among P&IDs, SCADA tags, and model states moved faster from proof-of-concept to trusted operation. A shared tag dictionary with provenance (i.e., units, timestamps, and transformations) allowed operators to trace a recommendation back to measured reality, reducing friction and easing change management. Where tags and model variables were ad hoc or duplicative, even well-calibrated models faced skepticism at the handoff to operations.

These mechanisms also clarify the observed differences between full- and pilot-scale results. Full-scale utilities, appropriately risk-averse and compliance-focused, tended to retain advisory-only mode unless constraint logic and governance were explicit that the “human in the loop” concept was maintained. Pilot environments, by design, bounded risk and latency, thereby enabling tighter loops and credible demonstrations of targeted actuation. Together, the cases suggest that realized value depends less on novel algorithms and more on whether the twin is architected to respect plant constraints, expose uncertainty, and interoperate cleanly with existing control layers.

The cross-case patterns explain why only a subset of surveyed systems meet strict DT criteria in practice (i.e., synchronized state with authorized write-backs). They also indicate a pragmatic path forward: encode constraints, right-size latency and uncertainty to the decision, and make the digital–physical mapping transparent enough for operators to trust.

8. Challenges, Limitations, and Current Approaches

The implementation of DTs in the water industry offers a promising pathway to improving operational efficiency, resource management, and sustainability. However, despite the considerable potential, there are several key challenges and limitations that hinder the full-scale adoption of DTs in the water sector. These challenges are often linked to the complexity of water systems, the need for real-time data integration, and the interaction between physical and virtual components. In many cases, existing infrastructures in the water industry are not adequately prepared to support the seamless integration of DT technologies.

Moreover, as DTs evolve to encompass more comprehensive functionalities, questions about data quality, computational demands, sensor accuracy, and workforce adaptability continue to emerge. Addressing these challenges is essential to unlocking the full potential of DTs and ensuring their applicability across diverse water management systems. The following sections outline the primary challenges that need to be addressed for the successful integration of DTs in the water industry.

8.1. Data Quality Assessment and Control

Ensuring high data quality is essential to the successful implementation of DT with its applications in the water industry. The accuracy and reliability of data directly influence the effectiveness of DTs, as poor data quality can lead to flawed simulations and suboptimal decision-making. Inaccuracies in measurement, inconsistencies in data transmission, and the inability of sensors to capture real-time changes can all compromise data quality. Therefore, maintaining data integrity is paramount to fully realizing the benefits of DTs, particularly in a sector where precision is critical for operational efficiency and sustainability.

One approach to enhancing data quality involves the use of reference measurements through manual inspection and maintenance. This includes regular calibration of sensors using hand-held references and controlled media. By routinely calibrating sensors against established standards, utilities can ensure that measurements remain accurate and consistent over time. This hands-on approach helps mitigate issues such as sensor drift or failures, which are common in long-term operations. However, this method can be labor-intensive and may not be scalable for large systems, where automated or semi-automated solutions might be more practical.

Another approach to address data quality challenges is defining typical ranges for key operational variables, such as flow rates or chemical concentrations. These ranges, based on historical data or engineering principles, can be used to flag abnormal values that may indicate sensor errors or system malfunctions. By identifying outliers early, this approach enables a more automated form of data verification and integrity maintenance, reducing the need for continuous manual monitoring.

A more robust approach to ensuring data quality is through redundancy in data collection. Redundancy can be implemented both spatially and temporally to validate sensor data and improve reliability. Spatial redundancy can be achieved through physics-based methods, such as mass balances, or empirical checks like the validation of typical ratios (e.g., ratio of total ammonia nitrogen (TAN) to total Kjeldahl nitrogen (TKN) < 1), which confirm sensor outputs by comparing them to expected norms. This ensures that data is cross-verified across different locations, providing a backup for when primary sensors fail or deliver inaccurate readings.

Temporal redundancy involves applying filters and smoothers to the collected data to detect and manage outliers and noises. For instance, placing upper and lower limits on expected values and their temporal derivatives helps maintain data consistency over time. Noise detection methods, such as monitoring changes in R² values, along with the use of low-pass filters, can further refine data accuracy, ensuring that the data is reliable for both short-term and long-term analyses.

Finally, a combination of spatial and temporal redundancy offers a comprehensive solution for ensuring data quality across an entire system. Techniques such as the application of Kalman filters, which combine spatial and temporal data for dynamic analysis, and nonlinear dynamic data reconciliation help synchronize and refine data from multiple sources. This combined approach allows for a continuous feedback loop between the physical and DTs, ensuring that real-time data flows are reliable, thus enabling more accurate simulations, predictions, and operational decisions.

Studies consistently report that data-quality limitations propagate into DT forecasts. At Eindhoven WRRF [35], nitrate and mixed liquor suspended solids (MLSS) predictions degrade during wet-weather conditions despite a robust data pipeline—illustrating how transient hydraulics and measurement uncertainty can impair calibration and short-term forecasts. At the urban–hydrology boundary, experimental evidence coupled with refined stormwater management model (SWMM) modeling shows that cascading green-gray layouts can materially reduce peaks and volumes relative to distributed configurations, underscoring that hydrologic connectivity is a boundary condition stormwater/sewer twins must represent for waterlogging mitigation [104].

Beyond individual sensors, disciplined metadata, consistent SI units, synchronized timestamps, documented sensor placement and transmission mode, and known error sources such as calibration drift, fouling, and scaling, is essential to assess fitness-for-purpose and troubleshoot anomalies. As outlined earlier in this review, establishing a “ground-truth” mapping across P&ID, SCADA tag names, laboratory identifiers, and model variables reduces integrity errors and commissioning time.

Mismatches between P&ID labels, SCADA tags, lab identifiers, and model notation remain a recurring source of integrity errors; several works in Section 7 highlight that this mapping is still a manual, error-prone step that benefits from standardized dictionaries or ontology-based approaches (see Section 7 cases: Changi [29], Bresso-Niguarda [98]).

8.2. Lack of a Framework for Developing Practical DT in the Water Industry

One of the major hurdles that the water industry faces in adopting DTs is the absence of a standardized, practical framework. Without a clear framework to guide the development, integration, and scaling of DTs, it is difficult for water utilities to implement these technologies on a large scale.

Modular frameworks offer a flexible solution to this problem. By developing DT architectures that can be modular and adaptable, water utilities can customize the digital infrastructure to fit their specific operational requirements. Modular frameworks allow different components of a DT to be integrated gradually, based on the needs and resources of each facility. This stepwise approach enables utilities of varying sizes to implement DTs without the need for extensive overhauls of existing systems, making the technology more accessible and scalable.

Regulatory alignment is another critical aspect that needs to be addressed. Engaging with regulatory bodies to develop compliance standards and benchmarks for DT performance is essential for ensuring that DTs meet the necessary safety, environmental, and operational requirements. Regulations around data security, environmental impacts, and system safety must be clearly defined, so that utilities have a framework within which they can confidently develop and deploy DTs. Aligning DT developments with existing and future regulations will also help streamline adoption by ensuring that the technology is compatible with legal requirements and best practices.

Collaborative platforms play a vital role in advancing DT technology in the water industry. By fostering collaboration between water utilities, technology providers, and research institutions, the industry can accelerate innovation and overcome challenges more effectively. Shared platforms enable stakeholders to exchange best practices, lessons learned, and technical innovations. This collective approach helps avoid redundancy in research efforts and supports the development of solutions that are broadly applicable across the industry. Collaborative platforms also encourage the development of interoperable systems, ensuring that DT technologies can work seamlessly across different facilities and regions.

8.3. High Cost of Sensor Implementation and Computational Resources

Although many variables are measurable, some operationally critical analytes remain costly to measure continuously and require frequent maintenance (e.g., calibration drift, fouling, scaling). These constraints are widely recognized in full-scale deployments and are reflected in DT-ready metadata guidance and plant experience (see Appendix C; Bresso-Niguarda [98]).

To reduce the sensing burden, soft sensors and hybrid models (mechanistic + machine learning) are increasingly used. In collection networks, theory-guided NNs with weak-form constraints improve drainage-pipeline siltation diagnosis from multivariate sequences and are strong candidates for soft sensors embedded in DT decision support during wet-weather transients [105]. In the reviewed cases, integrating virtual sensors and residual learning with mechanistic simulators improved the robustness of ammonia/nitrate predictions and enabled energy-aware control. At Cuxhaven (EWE Wasser GmbH), this approach supported aeration optimization with an annual power saving of approximately 1.2 million kWh while maintaining effluent quality [40].

On computing architecture, pragmatic historian-plus-edge pipelines can deliver frequent updates adequate for advisory decision support. Plant-wide twins such as Changi [29] and Eindhoven [35] illustrate the data and compute footprint—hundreds to more than a thousand tags—that must be planned for when considering cloud/edge trade-offs.

8.4. Lack of Automated Bidirectional Data Transfer

A fundamental aspect of DT is the bidirectional flow of data between the physical system and its digital counterpart. This two-way communication allows the digital entity to not only receive real-time data from the physical entity but also to autonomously send feedback, make adjustments, and optimize operations in the physical system. However, as highlighted in several studies reviewed in this study, many DT applications in the water industry face a critical limitation: the lack of automated data transfer from the digital to the physical entity. In these cases, the data flow remains unidirectional, where real-time data is transmitted to the digital model, but operational adjustments based on the model’s insights are either manually implemented or restricted entirely. Notably, most of the full-scale deployments we catalogued still operate in advisory-only mode, with only a minority enabling closed-loop write-backs (e.g., Lushan [34]), highlighting how governance and constraint-encoding remain practical bottlenecks.

This limitation undermines one of the core principles of a fully functional DT. While DS—where data flows in one direction—can be useful for monitoring purposes, they do not allow for real-time control and optimization of water systems, which is a key advantage of DT technology. The primary cause of this issue often lies in the stringent safety regulations and security protocols governing water utilities.

To address this challenge, several approaches can be considered:

• Hybrid Approaches with Manual Verification: One potential solution is the adoption of hybrid systems, where the DT generates operational recommendations based on real-time data, but these recommendations are subject to manual verification before implementation. This approach balances the need for safety and human oversight with the benefits of automation. Operators can review suggested changes from the DT and either approve or reject them based on additional safety checks or contextual factors that the DT may not fully account for.
• Incremental Automation: Another approach is to gradually increase the level of automation within water systems. Starting with low-risk operations, such as adjustments to operational variables (e.g., flow rates that fluctuate in response to demand) or controllable parameters (e.g., chemical dosing setpoints), can build trust in the system’s ability to operate safely. Over time, as the reliability of the DT is demonstrated, utilities may become more comfortable, allowing DTs to handle higher-level, more critical operations. This incremental approach ensures that any potential risks are managed while allowing the system to evolve toward full automation.
• Enhanced Cybersecurity Measures: One of the primary concerns for water utilities in enabling bidirectional data flow is the potential vulnerability to cyberattacks. Therefore, implementing robust cybersecurity protocols is essential for protecting both the physical and digital entities. Encryption of data, secure communication channels, and stringent access controls are necessary to ensure that the DT’s instructions are both safe and reliable. By mitigating security risks, water utilities may be more inclined to allow autonomous adjustments based on real-time feedback from the DT.
• Regulatory Collaboration: To facilitate automated data transfer, water utilities must work closely with regulatory bodies to ensure compliance with safety and environmental standards. Clear guidelines and frameworks that define the operational boundaries of DTs can help alleviate concerns around safety and security. Regulatory agencies can develop policies that outline how and under what circumstances DTs can make automated changes to water systems, ensuring that all operations remain within acceptable risk thresholds.

In summary, while the lack of automated bidirectional data transfer poses a significant challenge to the full implementation of DTs in the water industry, there are several approaches that can help overcome this barrier. By employing hybrid approaches, gradually introducing automation, enhancing cybersecurity, and collaborating with regulatory bodies, water utilities can begin to realize the full potential of DTs while maintaining the safety and security of their operations.

8.5. The Challenge of Control Authority

Throughout this review paper we described the structure and implementation of various forms of DTs in the water sector. The priority of water process facilities remains nonetheless to guarantee quality and quantity of the treated effluent, and digital aids such as DT are in support of the operators actions but cannot replace them. In fact, in most jurisdictions only licensed operators retain the control authority over the process equipment, de facto limiting the bidirectionality of DT to research or piloting efforts. Even in the case when operators would allow the DT to run the facility, the ability to override and take control of operations will always stem from the fact that operators and not computers are liable for the process actions and their consequences. This notwithstanding, the advent of digital tools in support of more informed control can only benefit a still operator-driven control and ultimately provide overlays of information (such as process unit costs, energy/power intensity of process, greenhouse gas intensity of treatment, etc.) that a normal SCADA system falls short of providing.

8.6. Operational Realities for Closing the Loop

Across the full-scale cases in this review, advisory-only operation is the norm. Even when a communication stack exists, automation is limited by governance and by how completely plant safety logic and operating limits are encoded in the DT. This is a socio-technical bottleneck: utilities must protect equipment, people, and compliance. Until interlocks, rate limits, and override rules are faithfully mirrored, the safest position is to keep the DT in recommend-only mode.

The key enabler is constraint logic. Physical limits and safety/regulatory constraints embedded in PLC/SCADA often determine whether any recommended action can actually run. If that logic is opaque or hard to access, model-based recommendations either fail at runtime or get blocked upstream, undermining operator trust. In practice, closing the loop depends less on a sophisticated optimizer and more on exposing, validating, and reproducing the plant’s real constraint set in the DT’s decision layer.

A quieter but essential task is variable mapping. Reliable operation needs a clean thread from P&IDs to SCADA tags/lab identifiers and finally model variables. Teams that establish a single “ground-truth” dictionary (or lightweight ontology) shorten commissioning and reduce silent errors.

Model choice should match the actuation you truly have. Where safety, interpretability, and control authority matter, mechanistic or hybrid (mechanistic + ML) models are usually preferable. Purely data-driven surrogates are strongest for soft-sensing and short-horizon forecasting. The deciding factors are observability, validation under relevant disturbances, and whether actuators can realize the recommended moves within the plant’s constraint envelope, not ideology.

Finally, performance claims need an evidence hierarchy. Vendor press releases provide context, not evidence. Quantitative statements should trace to peer-reviewed studies or utility-audited reports, with clear baselines, uncertainty, and time frames. Anything else is illustrative and should be labeled as such. Together, these operational realities explain why advisory-only is common today and point to a practical path to trusted actuation: make constraints explicit, make mappings unambiguous, match latency to purpose, choose models aligned with your authority, and insist on verifiable evidence as DTs move from dashboards to decisions.

9. Future Direction and Summary

This review has examined the deployment of DTs in the water sector, spanning both treatment and reuse applications. To date, most operational DTs remain advisory-only or are confined to specific subsystems; where they have been implemented, tangible benefits include energy savings, multi-day event forecasting, and enhanced operator decision support, albeit moderated by persistent data integration and standardization challenges (e.g., Changi [29], Eindhoven [35], EWE [40]).

Looking ahead, realizing the transformative potential of DTs in water management hinges on addressing several critical challenges:

• Full-Scale Demonstrations. While pilot and subsystem implementations have illustrated DT value, there is an urgent need for end-to-end, full-scale deployments across diverse water infrastructures—treatment plants, distribution networks, and even consumer-side applications. Such real-world examples will shed light on scalability hurdles and inform best practices for translating DT capabilities into operational resilience.
• Advanced Analytics Integration. The incorporation of sophisticated data analytics, particularly ML and AI, will unlock new levels of DT performance. By embedding adaptive learning algorithms, DTs can evolve alongside system dynamics, enabling continuous optimization of operations and more reliable forecasting under both typical and extreme conditions.
• Standardized Development Frameworks. Widespread adoption demands harmonized guidelines for DT creation and maintenance. Standardization efforts should cover data schemas, metadata conventions, sensor deployment strategies, and model validation protocols. Establishing these frameworks will facilitate interoperability, reduce implementation complexity, and ensure consistent performance across the sector.
• Sustainability and Resource Efficiency. As water scarcity intensifies and energy costs rise, DTs will be instrumental in steering systems toward sustainable operation. By coupling robust data pipelines with operator-focused use cases, such as energy-aware control and demand forecasting, DTs can deliver both environmental and economic dividends. For instance, Changi’s FDT achieved up to five-day event forecasting accuracy [29], while EWE’s ODT yielded approximately 1.2 million kWh in annual aeration energy savings [40].

However, these successes also expose areas for improvement. Accuracy degradations during wet-weather extremes at Eindhoven underscore the need for integrating hydrodynamic coupling and rigorous uncertainty quantification into DT calibration and forecasting workflows [35]. Furthermore, the scale of plant-wide integrations—tracking hundreds to over a thousand tags at Changi and Eindhoven—highlights the critical importance of metadata discipline and proactive maintenance planning.

To bridge research, policy, and operations, a short staged agenda is warranted: (1) strengthen uncertainty-aware estimation and calibration that remain robust in wet-weather conditions; (2) standardize DT-ready metadata, open APIs, and auditable change management to allow limited write-back; and (3) move from advisory-only to narrowly authorized setpoint changes, with PLC/SCADA constraint logic mirrored in the twin. In parallel, maintain a single plant tag dictionary (P&ID–SCADA–model) and align data and compute latencies with process time scales. Together, these measures convert priorities into near-term implementation and provide a controlled pathway from dashboards to dependable closed-loop operation.

Complementarily, mature multi-objective optimizers for large water-distribution resilience under cost constraints and cascade-reservoir scheduling can be coupled to synchronized twins (state estimation + constraint mirroring) to enable auditable, bounded write-backs at system scale [106,107].

In conclusion, DTs hold significant promise for revolutionizing water systems management. Realizing this vision will require coordinated efforts among researchers, industry practitioners, and policymakers to develop robust, scalable, and adaptable DT solutions. By addressing existing barriers—through full-scale demonstrations, advanced analytics, standardization, and sustainability-focused design—DTs can drive the next generation of efficient, resilient, and sustainable water infrastructure worldwide.