A Systematic Literature Review on Flow Data-Based Techniques for Automated Leak Management in Water Distribution Systems

Highlights

What are the main findings?

• IoT, smart metering, and AI-based models are increasingly used for real-time leak management, but their effectiveness relies on data quality and system integration.
• While significant research focuses on leak detection algorithms, fewer studies address the full scope of automated leak management systems, limiting progress beyond detection.

What is the implication of the main finding?

• Developing an automated leak management system based on advanced data acquisition, robust leak management models, and scalable real-time monitoring platforms is crucial for enhancing leak detection accuracy and responsiveness.
• Further research is essential to improve model accuracy, scalability, and system integration, addressing key challenges for fully automated leak management deployment.

Abstract

Smart cities integrate advanced technologies, data-driven decision-making, and interconnected infrastructure to enhance urban living and resource efficiency. Among these, Smart Water Management (SWM) is crucial for optimizing water distribution and reducing Non-Revenue Water (NRW) losses, a persistent challenge for utilities worldwide. Water leaks contribute significantly to NRW, necessitating real-time leak detection and management systems to minimize detection time and human effort. Achieving this requires seamless integration of SWM technologies, including advanced metering infrastructure, the Internet of Things (IoT), and Artificial Intelligence (AI). While previous studies have explored various leak detection techniques, many lack a focused analysis of real-time data integration and automated alerts in SWM systems. This Systematic Literature Review (SLR) addresses this gap by examining advancements in automatic data collection, leak detection models, and real-time alert mechanisms. The findings highlight the growing potential of data-driven approaches to enhance leak detection accuracy and efficiency, particularly those leveraging flow and pressure data. Despite advancements, model accuracy, scalability, and real-world applicability remain. This review provides critical insights for future research, guiding the development of automated, AI-driven leak management systems to improve water distribution, minimize losses, and enhance sustainability in smart cities.

1. Introduction

The concept of smart cities revolves around integrating advanced technologies, data-driven decision-making, and interconnected infrastructure to enhance urban living. Smart cities encompass various domains, such as smart buildings, smart transportation, smart governance, and smart environments, all aimed at improving the efficiency and sustainability of urban services. They rely on cutting-edge technologies and sustainable clean energy innovations supporting urban resilience and environmental goals. Recent studies have shown a crucial link between energy sustainability and smart urban ecosystems [1,2]. Among these components, Smart Water Management (SWM) ensures the efficient utilization and distribution of water resources. SWM integrates Internet of Things (IoT) technologies, real-time sensor networks, big data analytics, and Artificial Intelligence (AI) to optimize water distribution, monitor consumption patterns, and improve system resilience.

One of the most critical challenges in SWM is managing Non-Revenue Water (NRW), representing the water volume lost before reaching consumers. NRW accounts for approximately 30% of the total water supply globally, leading to an annual estimated economic loss of $14 billion [3]. The sources of NRW can be categorized into apparent losses caused by metering inaccuracies or unauthorized consumption and real losses primarily resulting from leaks in transmission lines, distribution mains, service connections, and storage tanks [4]. Beyond economic consequences, water leaks pose serious environmental and public health risks, exacerbating water scarcity and increasing infrastructure maintenance costs [5,6]. Addressing these challenges requires effective leak detection and management strategies to ensure the sustainability and resilience of Water Distribution Systems (WDS).

An efficient leak management system should detect leaks accurately and minimize detection time through real-time monitoring and alerts. This can be accomplished by implementing an automated system grounded in Smart Water Management (SWM) principles. However, despite its potential, challenges remain in integrating diverse data sources, maintaining high data quality, and addressing infrastructure limitations. Collecting and processing high-frequency data in real-time are key requirements for automated leak detection. Technologies such as smart water meters, Automatic Meter Reading (AMR), and Automatic Meter Infrastructure (AMI) facilitate automatic data collection, enabling the continuous transmission of high-frequency data essential for leak detection. These technologies rely on various sensors, including flow, pressure, and acoustic.

Flow sensors have proven particularly effective for leak detection [7,8,9,10]. Unlike pressure-based systems, which can be limited by transmission distance, signal strength, and sensor costs [11], flow sensors provide continuous monitoring, aiding in leak detection and infrastructure planning, water distribution optimization, and demand forecasting. As a result, flow-based techniques present a practical and cost-effective solution for automated leak management.

Over the years, various methodologies have been proposed for leak detection and management. Several studies have reviewed these methods from different perspectives. Puust et al. [12] classified leak management into assessment, detection, and control measures. Hu et al. [13] examined the data-driven and model-based methods for leak detection. A transient-based leak detection technique that uses pressure signals for water leakage detection was studied by Colombo et al. [14]. Li et al. [15] categorized water leak detection techniques into hardware- and software-based approaches. External device-based processes and internal device-based approaches in leak detection were reviewed by Adedeji et al. [16]. However, most reviews focus on leak detection algorithms while neglecting other critical aspects of leak management, such as leak localization, assessment, and real-time alert systems. Moreover, these reviews commonly include studies based on heterogeneous data sources, ranging from low-resolution datasets (water bills, interviews, and surveys) to high-resolution water meter data, many of which lack real-time capabilities. In contrast, this review exclusively focuses on studies that utilize real-time water flow data to support the development of automated leak management systems grounded in SWM principles.

This systematic literature review aims to comprehensively understand leak management techniques in WDS to address these gaps, focusing on flow data-based methods for automated leak management. While this review covers various leak management strategies, from detection to localization and control measures, it emphasizes evaluating their effectiveness for automation. This review leverages IoT-enabled, real-time monitoring to explore how these methods can be integrated into automated systems. Additionally, this review assesses existing implementations of automatic leak detection systems, identifying notable advancements and persistent issues. The findings highlight the growing adoption of automatic data collection and real-time visualization systems in leak management. Furthermore, there is an increasing reliance on AI-based models for real-time leak detection and management, demonstrating their potential to improve the efficiency and effectiveness of leak management systems significantly.

The remainder of this paper is organized as follows: Section 2 outlines the research methodology used in this review; Section 3 presents a comparative analysis of automated leak management workflows; Section 4 discusses key research questions, focusing on flow data applications and processing techniques for automatic leak detection; and finally, Section 5 summarizes the findings and outlines directions for future research.

2. Methodology

A systematic review of the literature was performed in line with the guidelines for systematic review in systems and automatic engineering [17] and following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses-2020 (PRISMA) statement [18]. This review was performed in three phases: planning, conducting, and reporting the review. Research questions were formulated based on the Problem, Constraints, and Systems (PCS) framework [19] during the planning phase, and review protocols were developed. In the second stage, a manual search process in various research and academic databases was performed, and eligible publications were manually inspected to extract relevant evidence for analysis. The final stage included the documentation of the work. The following sections discuss each of these stages in detail.

2.1. Planning the Review

This study aims to address the following research questions.

RQ1 What are the existing methods of leak management using water flow data?

• Rationale: This question identifies and categorizes current techniques and methodologies that use flow data for leak management.

RQ2 How do the different methods addressed in RQ1 apply to automated leak management?

• Rationale: This question evaluates each method’s level of automation by comparing factors such as accuracy, real-time response, visualization, and localization, assessing how well each supports fully automated leak management.

RQ3 What is the strength of the evidence in support of the different methods?

• Rationale: RQ3 addresses the quality and reliability of research supporting each method, focusing on experimental design and real-world testing.

RQ4 What implications will these findings have when creating a real-time automatic leak management system?

• Rationale: This question seeks to determine how insights from comparing methods and assessing evidence can guide the practical development and integration of an effective, real-time leak management system utilizing flow data.

The review protocol used in this study outlines the methodology used to systematically identify, select, and analyze research on water leak management using flow meter data. This review aims to analyze the methodologies and technologies in applying leak detection systems to provide insights into advancements and research gaps in the field. To ensure a comprehensive literature review, a structured search strategy was implemented. Relevant journal articles and conference proceedings were retrieved from selected academic databases. The search was conducted using both primary and alternative keywords to maximize coverage. The primary keywords included “water leak management”, “smart water meter”, and “flow data”. In addition, alternative keywords such as “burst detection”, “anomaly detection”, “anomaly assessment”, and “water loss” were used for water leak management. At the same time, “automated meter readings”, “intelligent meters”, “flow meters”, “automated meter infrastructure”, and “data-driven” approaches were considered for smart water meters to represent an automated data collection system.

A single-line search strategy utilizing Boolean operators was applied to streamline the search process and ensure alignment with the research scope. The following search strings were used:

((“water”) AND (“leak” OR “leakage” “burst” OR “anomaly” OR “loss”) AND (“detection” OR “assessment” OR “management”) AND (“smart meter” OR “intelligent meter” OR “flow meter” OR “data-driven” OR “Automated meter Infrastructure”)).

This review utilized several well-established academic databases, as shown in

Table 1. Summary of databases used for review.

Database	Selected Records	Search Field
American Society of Civil Engineers	25	All Content
Civil Engineering Database	14	All Fields
Engineering Village (Compendex, Geobase, Inspect)	169	Abstract, Title, Keyword
ScienceDirect	30	Abstract, Title, Keyword
SciTech Premium Collections	21	Peer reviewed
Scopus	36	Abstract, Title, Keyword
Springer Online	14	All content
Taylor and Francis	14	Abstract
Web of Science	50	Abstract, Title, Keyword

. These databases have rigorous indexing standards and primarily include peer-reviewed journals and high-quality conference proceedings, thereby ensuring the reliability and academic quality of the review. In addition to selecting papers from reputable sources, each study’s quality was assessed based on the clarity of methodology, use of real-world data, and the relevance of the findings to automated leak management. Specific inclusion and exclusion criteria were applied to maintain the selected studies’ relevance and quality. This study had to focus on water leak management and include at least one data source based on flow meter data. Studies relying solely on pressure, vibration, or acoustic sensors were excluded. Furthermore, only research utilizing real-life flow data was considered, while studies using software-simulated data for both training and testing were excluded. This review was limited to works published in English, and review papers were omitted to ensure the inclusion of original research and avoid redundancy.

Since this review aimed not to compare individual studies but to understand the extent of development in automated leak management systems, peer-reviewed journal articles and relevant conference proceedings were included. This approach allowed for the capture of a comprehensive picture of current and emerging trends in the field, including early-stage innovations and real-world applications.

2.2. Conducting the Review

The articles were retrieved based on an academic search, snowballing, and grey literature capture. The publication details and abstracts were downloaded, and MS Excel was used to tabulate the information for a quick overview. Reference manager software such as Mendeley (v2.131.0) was used to manage the collected data.

Initially, academic literature searches were performed on different research databases, as shown in Table 1. Figure 1 shows the PRISMA 2020 flow diagram for systematic reviews summarizing the selection process. A total of 373 records were retrieved based on the academic database search. The Abstract, Title, and Keyword have been used as a search field in most of the databases that support this advanced search option; for the others, search based on Abstract, all contents, or Peer Review has been employed. About 207 records remained after removing the duplicates using the Mendeley reference manager software. In addition, 30 records were retrieved based on snowballing.

The records obtained from the academic literature search were considered primary studies and were analyzed based on the title, keywords, and abstract. The studies were evaluated based on quality and the defined inclusion and exclusion criteria for the preliminary selection. Of the 207 selected records, 140 were initially screened, and 132 full papers were retrieved from various databases. These included 98 journal articles and 34 conference papers.

The next step in the selection process was reading the full document to make the final selection. The accessed reports were further checked for eligibility, among which eight reports were removed based on quality checks, and 71 records identified via database searches were removed based on the inclusion and exclusion criteria. Similarly, among the reports accessed through snowballing and grey literature searches, 20 were removed based on the eligibility criteria as per Section 2.1. Thus, a total of 71 reports were included for the review.

2.3. Reporting the Review

The final step in performing the systematic literature review was the reporting of the review by documenting the extracted data. The contents were organized based on the PRISMA 2020 checklist. The result of the systematic review is discussed in the following section.

3. Results

The descriptive statistics of the literature review based on the final selection of studies are discussed in this section. Of the 71 selected records, 11 are conference proceedings, and 60 are journal articles from various journals, as shown in Figure 2. These journals cover water resources, environmental management, sustainability, engineering, pipeline systems, and advanced technology applications. They focus on key areas such as water infrastructure, hydrology, resource management, sustainable practices, and integrated water management, emphasizing the role of information technology, data analysis, and modeling. This diversity enables a comprehensive exploration of topics related to automatic leak management systems.

Figure 3 shows the distribution of the selected articles by year. There is a noticeable increase in publications from 2014 to 2024 that utilize smart meter technology. Although many papers on leak detection were published earlier, most relied on manual meters or other data sources, such as interviews, questionnaires, and utility bills, which are outside the scope of this study. The upward trend in publications over the years highlights the growing potential of automatic data collection techniques in the water utility sector.

3.1. Automatic Leak Management System Overview

An automatic leak management system is designed to continuously monitor, detect, and respond to water leaks in a distribution network with minimal human intervention. It integrates real-time leak detection, AI-based leak detection algorithms, automated alert mechanisms, and advanced visualization tools that enable timely response and precise control. From the reviewed literature, three key components emerge as essential for effective leak management: a data acquisition and management system that collects, transmits, and stores information from sensors and meters; a leak detection and prediction model that processes this data using analytical or AI-based methods; and a leak monitoring platform that functions as the human–machine interface for online monitoring, leak localization, and alert visualization. Figure 4 illustrates the integrated data pipeline of an automatic leak management system, synthesized from the reviewed studies.

The literature review reveals that only 35% of the studies address the development of a comprehensive leak management system encompassing all three subsystems. In contrast, most studies focus on developing leak models that could be integrated into such systems. The following sections dwell on the review results corresponding to each subsystem in automatic leak management.

3.2. Data Acquisition and Management

The initial phase of an automatic leak management system involves acquiring hydraulic data. Figure 4 presents the system architecture for an automated flow data acquisition setup, comprising data collection, transmission, and storage components. This review focuses on leak detection based on flow measurements, with flow sensors and flow meters serving as the primary data sources. Smart water meters equipped with integrated flow sensors are also commonly used.

Flow data collected from sensors is transmitted to a central server for analysis using wireless communication technologies, such as LoRaWAN [20], NB-IoT [21], GSM-GPRS modules [7,22,23], or mobile networks [24]. Data transmission typically utilizes protocols like MQTT, HTTP, FTP, SMS, or APIs. In some setups, sensors are connected to edge devices such as Raspberry Pi, which perform local data processing and can trigger immediate alerts on-site. Depending on the system configuration, this locally processed data may be stored locally or transmitted to a central server for further processing and integration.

Data transmission frequencies vary from daily updates to intervals as short as one minute. The acquired flow data are stored as time series in databases or cloud platforms, with sampling rates typically ranging from one second to one hour. The most used intervals are five minutes (19%), fifteen minutes (27%), and thirty minutes (19%) (Figure 5). A variety of databases and cloud platforms are employed across different studies, including MS Access [22,25], MySQL [26,27], PostgreSQL [24,28,29], and Amazon Web Services (AWS) [30].

Various automatic data collection technologies leveraging flow sensors are reported in the literature, including flow meters with data loggers, smart water meters, AMR, AMI, and Supervisory Control and Data Acquisition (SCADA) systems. However, approximately 30% of the reviewed studies mention using flow sensors or flow meters as data measurement tools without further detailing the system architecture. Figure 6 illustrates the trend in automatic data acquisition methods obtained from the selected literature over three periods: before 2011, between 2011 and 2017, and after 2017. Traditional methods like water meters with data loggers dominated before 2011 but declined sharply, with minimal use after 2017. In contrast, systems like AMR and SCADA gained traction and remained relevant but with less dominance. Significant growth is observed in smart water meters and AMI in the 2018–2024, reflecting the increasing adoption of advanced, real-time, and integrated systems for water management.

3.2.1. Water Meters with Data Loggers

Flow meters with data loggers were widely used in earlier studies to collect and store flow data. Various types of flow meters, such as pulse water meters and pulse emitters, were employed for flow measurement [31]. The data loggers record and store water consumption data locally and transmit it to a central database at scheduled times for further analysis using different connectivity options. These connectivity options includes General Packet Radio Service (GPRS) [7,25,32], Global System for Mobile Communications (GSM) [33], and IoT technologies like Narrowband IoT [21].

3.2.2. Automatic Meter Reading

AMR technology enables utilities to collect water consumption data directly from water meters without manual intervention. Flow sensors, such as impulse [24,34] and Hall effect sensors [26], are commonly used in this technology. The water meters have an encoding system connecting to an endpoint to capture flow data. Radiofrequency transmitters transmit this data from the flow sensors to the receivers. The receivers gather the data under direct supervision and transmit it to local or central information systems or databases, such as SQL [26] or PostgreSQL [24]. While AMR is effective for billing and detailed analysis, it is unsuitable for real-time applications.

3.2.3. Supervisory Control and Data Acquisition System

A SCADA system is an integrated solution to monitor and control industrial processes and infrastructure. It enables automated, real-time data acquisition and remote management of assets. In the field of water management, SCADA systems are typically categorized into two types: SCADA Q, which measures source flows, and SCADA P, which tracks pipe flows and internal pressure in addition to source flows [31]. SCADA systems provide connectivity based on modern IT standards, supporting SQL and web-based applications. The flow data collected by SCADA systems have been used in various studies related to leak detection [31,35,36,37,38,39,40] and anomaly detection [41,42]. However, SCADA systems have limitations in leak localization and are less effective in detecting small leaks [31].

3.2.4. Smart Water Meters with IoT

A smart water metering system is designed to monitor and detect anomalies in a real-time water distribution network and efficiently prioritize and manage maintenance tasks. Microcontrollers embedded within the system process signals obtained from flow sensors, converting them into time-series data for further analysis [43]. The data are transmitted to end users using a wireless communication infrastructure at predefined intervals. Integrating technologies in smart water meters enables automated and real-time data acquisition and analysis. Smart water meters integrated with IoT technologies, such as Wi-Fi and 3G/4G modules [30,44,45], offer advanced online data acquisition capabilities. These technologies enable real-time transmission of water consumption and flow data to centralized or cloud-based systems, facilitating continuous monitoring, analysis, and management of critical flows in water distribution networks. Additionally, low-power wide-area networks, such as LoRaWAN, have gained traction in this domain for their efficiency and extended range [20]. Commercially available wireless and mobile communication technologies, such as GSM and GPRS, are also widely utilized [22,23,46,47].

3.2.5. Automatic Meter Infrastructure

An AMI system transmits data directly to utilities at regular intervals without human intervention. The key components of an AMI system include smart water meters, communication networks, and meter data acquisition and management systems. IoT technologies form the backbone of AMI systems, enabling seamless data flow through interconnected devices and sensors deployed across the water distribution system. Unlike AMR, which supports only unidirectional communication, AMI enables two-way communication, allowing for real-time monitoring and control. IoT-based AMI architecture enables smart meters to interact with utility management systems continuously. AMI systems provide real-time data on end-user demand, which can be stored and used to model a WDS accurately [48]. Studies have also demonstrated that an IoT-enhanced AMI-based collection system enhances leak detection accuracy and reduces false alarms [31].

3.3. Leak Management Model

The next phase in an automatic leak management system involves developing a leak management model. This process typically includes preprocessing the collected data, designing statistical or machine learning-based algorithms for leak management, and evaluating their performance. Puust et al. [12] state that leak management can be broadly categorized into leak assessment, detection, and control measures. Studies have also focused on leak localization and leak prediction as part of comprehensive management approaches.

Figure 7 summarizes the various leak management techniques identified in the reviewed literature, highlighting that most studies (58%) focus only on leak detection, which involves identifying the presence and approximate location of water leaks. Leak detection systems generate alerts when leaks occur, but often do not quantify their severity or impact. Leak assessment, on the other hand, aims to evaluate the characteristics of leaks, including their severity and the volume of water lost. Unlike detection, leak assessment does not typically consider the spatial location of leaks. 6% of the reviewed studies exclusively focus on leak assessment, while 9% combine leak detection with assessment. A smaller but notable proportion (10%) combines detection with localization to identify precise leak locations, while another 10% focuses on anomaly detection, capturing abnormal events such as leaks. An anomaly can occur due to increased water consumption, leaks/bursts, and system failure. Anomaly detection is usually followed by system event classification. Only studies explicitly addressing burst/leak detection as an anomaly are included in this review. More advanced approaches, such as leak prediction and standalone leak localization, are less explored, constituting just 1% of the studies. Only 3% of the studies consider leak detection, leak size estimation, and leak localization together. Finally, 4% of the research addresses other aspects of leak management, reflecting a diverse but predominantly detection-oriented body of work.

3.3.1. Data Analysis Techniques

Figure 8 illustrates the distribution of various algorithms and approaches used for leak management in the reviewed studies. Leak assessment methods are categorized into top-down, bottom-up, and hybrid approaches, with hybrid methods combining elements of both. These approaches have been employed for both leak detection and assessment when integrated with data from multiple water meters.

A small proportion (5.2%) of those focused on leak assessment or combined leak detection and assessment employ top-down modeling approaches. These primarily use the water balance approach that calculates the water balance within a DMA using time-series inflow, outflow, and consumption data. While the traditional water balance approach relies on smart water meter readings, modifications incorporate parameters such as apparent loss coefficients, leakage area, and unmeasured authorized flow rates for improved leak assessment [40].

The bottom-up approach, such as the Minimum Night Flow (MNF) method, is used in 9.09% of the total studies. It utilizes flow data recorded during late-night and early-morning hours, when user activity is minimal, to detect leaks. Elevated MNF values may indicate leaks, with Legitimate Nighttime Consumption (LNC) factored into the analysis to account for reasonable water usage during this period [49]. Additionally, 6.49% of studies employ hybrid methods that combine MNF and water balance techniques for enhanced leak detection and assessment.

Leak detection studies encompass supervised, semi-supervised, and unsupervised techniques, as detailed in

Table 2. Synthesis of existing leak and anomaly detection techniques.

Approach		Author	Algorithm		Data
Supervised	Data Driven Classification	[50]	Variational Autoencoder (VAE) and Support Vector Machine (SVM)		Flow
		[35]	1D Convolutional Neural Network (CNN)		Flow, Pressure, Temperature
		[51]	Multi-Layered Perception (MLP)-Artificial Neural Network (ANN), Time Delayed Neural Network ANN		Flow, Pressure, Temperature
		[52]	Random Interval Spectral Ensemble (RISE)		Flow
		[36]	Random Forest (RF)		Flow
	Model Based Classification	[20]	K-Nearest Neighbors, Linear SVM, Radial Basis Function SVM, Decision Tree (DT), RF, AdaBoost, Gaussian Naïve Bayes (GNB)		Flow, Pressure
		[48]	2D CNN		Flow, Pressure
		[34]	Artificial Neural Network (ANN) and SVM Classification		Flow
		[53]	DT, RF, SVM, Logistic Regression (LR), GNB, MLP		Flow, Pressure
	Prediction Classification		Prediction	Classification
		[25]	Mixture density model (MDN)	Fuzzy classification	Flow, Pressure
		[54]	Symbolic regression	K-sigma method (Std deviation)	Flow
		[55]	Long Short-Term Memory (LSTM)	Multi-thresholding classification	Flow, Pressure
		[56]	Bayesian dynamic linear model	Thresholding	Flow, Pressure
		[9]	Weighted least squares-based expectation maximization (EM)	-Thresholding	Flow
		[22]	Adaptive Kalman Filtering	Thresholding	Flow, Pressure
		[32]	MDN-ANN	Fuzzy Inference System	Flow, Pressure
		[57]	Adaptive water demand forecast	Thresholding	Flow, Pressure
		[58]	Recurrent Neural Network (RNN)-LSTM	Multi-thresholding (X chart method)	Flow
		[39]	Empirical Mode Decomposition (EMD) and hydraulic modeling	3-sigma Statistical Process Control (SPC) and sliding window-based SPC	Flow, Pressure
		[59]	RNN-LSTM	Multi-threshold classification	Flow
		[60]	ANN	SPC	Flow, Pressure
		[61]	MDN	Rule-based model	Flow, Pressure
		[62]	MDN-ANN	Classification module based on time window	Flow, Pressure
		[10]	Kalman Filtering	Estimation of residuals	Flow, Pressure
		[63]	Adaptive Kalman Filtering	Thresholding	Flpw
		[64]	Multifactor XGBoost model, RF, ANN	CNN	Flow
		[7]	Support Vector Regression (SVR)	Thresholding (daily mean consumption)	Flow, Pressure
		[65]	Seasonal ARIMA and Transfer function	95% Confidence Interval	Flow
		[41]	Support Vector Regression (SVR)	Mean Absolute Percentage Error (MAPE)	Flow
		[66]	ML-Extended Kalman Filtering		Flow, Pressure
Unsupervised	Statistical approach		SPC-Exponentially Weighted Moving Average (EWMA), Xbar
		[27]	Statistical thresholding		Flow
		[26]	Statistical thresholding		Flow
		[21]	Statistical thresholding		Flow
		[44]	Statistical thresholding		Flow
		[67]	Statistical thresholding		Flow
		[45]	Statistical thresholding		Flow
		[68]	Correlation		Flow
		[69]	SPC-Western Electric Company rules (WECO), Cumulative Sum Control Chart (CUSUM), EWMA, Hotelling T2, Multivariate Cumulative Sum (MCUSUM), Multivariate Exponentially Weighted Moving Average (MEWMA)		Flow, Pressure
		[70]	SPC- EWMA		Flow, Pressure
		[31]	SPC-WECO		Flow
		[38]	SPC-Multivariate statistical approach		Flow
		[71]	SPC-X-bar, CUSUM, Seasonal Hybrid Extreme Studentized Deviate (SH-ESD)		Flow, Pressure
		[72]	SPC-CUSUM, EWMA		Flow, Pressure
	Clustering and pattern matching	[73]	Shape similarity-based approach for pattern recognition		Flow
		[8]	Clustering based on cosine distance		Flow
		[74]	Self-Organizing Map (SOM)		Flow
		[75]	Shape similarity-based approach for pattern recognition		Flow
		[76]	Density-based clustering		Flow
		[42]	Outlier regions, clustering		Flow
		[77]	Binary Associative Neural Network		Flow
Semi-Supervised	Classification	[33]	EM algorithm		Flow, Pressure
Leak assessment-based methods	Bottom Up	[23]	MNF		Flow
		[30]	MNF		Flow
		[49]	MNF modeling based on Fixed and Variable Area Discharges (FAVAD) Principle
		[78]	MNF		Flow
		[79]	MNF-based change detection		Flow, Pressure
	Top Down and Bottom Up	[24]	Water Balance, MNF approach, Probability Density Function (PDF)		Flow
		[28]	Water Balance and Automated MNF		Flow
		[29]	Water Balance, MNF approach, PDF		Flow
		[46]	Water Balance and MNF		Flow
	Top Down	[40]	Water balance model based on Adaptive Moment Estimation (Adam)		Flow, Pressure
		[80]	Mass balance based on hydraulic modeling		Flow, Pressure
		[47]	Water Balance based on hydraulic modeling		Flow, Pressure

. This table supports the percentage values discussed in the text by linking them to the specific literature sources.

Supervised methods include both classification and prediction-classification-based approaches. Among the reviewed studies on leak management (Figure 8), 11.68% focus on supervised classification (data driven and model based), where machine learning models are trained using historical water meter data labeled with leak and non-leak scenarios to distinguish between these conditions. Advanced AL-based methods like Variational Autoencoders (VAEs) combined with SVMs are also employed for dimensionality reduction and improved accuracy [50]. Given the scarcity of labeled data for leaks, 40% of the supervised classification studies use model-based classification that relies on simulated datasets generated using tools like EPANET 2 [34], EPANET 2.2 [48,53], or tools such as LoRaSURFING [20]. Data-driven approaches based on self-supervised learning, where the model does not need external class labels and instead uses labels assigned to generated data, are also used [52].

The prediction-classification approach was utilized in 27.27% of the studies (including leak detection and anomaly detection) and involves building predictive models using water consumption data, excluding outliers, to forecast future usage. Potential leaks and anomalies are identified by detecting significant deviations between predicted and actual values, making this method effective even without historical leak data. Its success relies heavily on the accuracy of the predictive model, which is typically developed using advanced AI techniques like deep learning, while classification is commonly performed using simple thresholding methods (Table 2).

The unsupervised techniques include statistical approaches and clustering. The statistical approach used in 19.48% of the studies (leak and anomaly detection) includes simple thresholding, correlation analysis, and the Statistical Process Control (SPC) technique to detect leaks. The SPC identifies non-random patterns in water distribution systems using statistical methods like Shewhart control charts to detect mean shifts [69]. Univariate SPC methods (e.g., WECO, CUSUM, and EWMA) analyze individual parameters. In contrast, multivariate methods (e.g., Hotelling’s T², MCUSUM, PCA, and MEWMA) examine correlations among multiple parameters to create a single anomaly indicator [69,81]. Finally, 7.79% of the studies adopt clustering-based and pattern-matching techniques. Clustering techniques treat leak detection as an anomaly detection problem by clustering flow data and identifying outliers as leaks. Algorithms such as Rodriguez and Laio’s theory of Burst detection in district metering areas using a data-driven clustering algorithm [8,76] and Self-Organizing Maps (SOM) [74] are commonly used for this purpose. The pattern matching techniques compare the flow patterns against normal or leak patterns and flag the anomalies. A small percentage (1.3%) of the selected studies focus on semi-supervised classification, utilizing the expectation maximization algorithm with a hybrid mixture model combining Gaussian and t-distributions.

Leak localization methods are model-based (6.49%) and data-driven (3.90%). Most studies (75%) use pressure-dependent methods, where hydraulic models estimate nodal pressure based on water demand, and pressure drops are analyzed to identify leaks. The largest drop typically occurs at the leak location due to accumulated energy losses. Approaches like binary integer linear programming models [31], clustering [79], uncertainty optimization with weekly water consumption [39], and graph theory for valve operations [82] are used to enhance localization accuracy. Data-driven techniques use inputs like acceleration, pipe area, and time data [45] or machine learning models such as CNN for leak detection. Pressure gradient analysis across a DMA and seasonal differencing can produce three-dimensional pressure maps for precise leak localization. These methods rely on strategically placed pressure sensors within the network. Only a small fraction (1.3%) of the reviewed literature addresses leak prediction. For instance, McMillan et al. [50] proposed a hybrid deep learning framework called FLUIDS (Forecasting Leakage and Usual Flow Intelligently in Distribution Systems), which utilizes a deep learning-based RNN-LSTM approach for forecasting leaks at the district metered area level.

3.3.2. Data Preprocessing

The leak detection model’s accuracy depends on the data quality. Thus, data preprocessing is one of the most critical steps in managing smart water meter data. Some sensors may possess missing or false data due to operational problems in the communication system, such as between the data loggers and sensors [56] Hence, it becomes important to identify gaps or missing values and fill them with “Not A Number (NAN)” or remove the data altogether. Various preprocessing techniques have been used in the literature, including noise removal, data completion, data resampling, data normalization, data reconstruction, data reformatting, data decomposition, and data transformation. Some of the commonly used data preprocessing techniques are described in

Table 3. Data preprocessing techniques.

Data Preprocessing	Summary	Studies	Data Analysis Technique
Data cleaning/Noise removal	Filter values exceeding the threshold	[24]	Top Down and Bottom Up
	Apply Chebyshev’s inequality	[29]
	Dynamic time warping and thresholding	[36]	Supervised classification
	Filter using ARIMA	[51]
	Remove repeated measurements	[56]	Prediction classification
	Filter using predefined criteria	[61]
	Denoising using wavelets	[60]
	Check for data errors	[66]	SPC, Prediction classification
	Check for data errors	[71]	Statistical approach
	Remove based on correlation and distance measures	[68]
	Handle missing and duplicate data	[72]
	Remove using thresholds	[42]	Clustering
Data completion	Kalman smoothening for faulty data	[50]	Supervised classification
	Interpolating repetitive values	[59]	Prediction classification
	ARIMA filter to fill missing values	[61]
	Kalman smoothening	[83]	Leak prediction
Data resampling	Resampling to 1 h data	[56]	Prediction classification
	Resampling data into parallel flow sets	[9]
	Normalized to a common time step	[46]	Top down and bottom up
	Normalized to a common time step	[42]	Clustering
Data Normalization	Mean and Std. deviation	[8]	Clustering
		[75]
		[69]	Statistical approach
		[73]	Pattern matching
		[46]	Top down and bottom up
		[51]	Supervised classification
		[61]	Prediction classification
		[77]	Pattern matching
	Low-pass filtering to obtain smoother normalized forms of residual curves	[63]	Prediction classification
Data Reconstruction	Consumption statistics of each consumer	[46]	Top down and bottom up
Data Reformatting	Input stream into a tapped delay line format	[61]	Prediction classification
Data Decomposition	Data decomposition based on Seasonal Trend decomposition procedure based on Loess (STL)	[72]	Statistical approach
	Decompose the data to seasonal and non-seasonal components	[71]
Data Transformation	Converting time-series data to matrices based on time intervals	[58]	Prediction classification
		[38]	Pattern identification and SPC
		[76]	Clustering

.

The specific requirements of the data analysis technique determine the selection of preprocessing steps. Data denoising is crucial to eliminating noise in time-series data from flow meter data, which can otherwise result in false alarms. Additionally, denoising improves prediction accuracy in machine learning-based leak detection methods [36,51]. It plays a vital role in prediction classification and unsupervised approaches, such as statistical methods and clustering, where clean, outlier-free data are essential to establish a baseline model for leak detection through predicted readings. Data filtering techniques, which identify and exclude anomalies, are also employed to manage noise [24,29,51]. The Dynamic Time Warping (DTW) algorithm effectively filters abnormal data in pseudo-periodic consumption patterns [36].

Data completion techniques address missing or repetitive values in time-series datasets. Common methods include interpolation [59], ARIMA filtering [61], and Kalman smoothening [50,83]. These approaches ensure that incomplete datasets are reconstructed to maintain data continuity, particularly critical for data-driven analyses like MNF and water balance calculations. For instance, Loureiro et al. [46] proposed reconstructing missing data using consumption statistics at various levels, such as individual consumers, consumer groups, or the entire network.

Another important preprocessing technique is data resampling, which involves aggregating or interpolating data to align with different time intervals or frequencies. This step is often necessary for time-series analysis and modeling, particularly when datasets come from sources with varying temporal resolutions. Data normalization is also a frequently employed preprocessing step in studies. Since water consumption patterns are predominantly time-dependent and often exhibit diurnal variations, the hydraulic characteristic values, like mean and standard deviation, can vary considerably. Normalizing these values and aligning them with statistical principles becomes imperative for developing and applying effective leak management strategies. Normalization ensures that data are on a consistent scale, making comparing and analyzing different aspects of the WDS easier.

Data normalization is essential in statistical-based analysis, especially in statistical process control techniques [69,73] and clustering [8]. These methods assume that the data are stationary, making it essential to perform normalization. However, outliers can skew statistics like the mean and standard deviation. To address this, more robust measures have been adopted [70]. Water consumption patterns have varying statistical characteristics due to regular human behavior. Data at different times of day and different days of the week follow different distributions. The accuracy of methods such as SPC could be highly affected if applied directly to raw monitoring data. Thus, identification and removal of periodic trends and seasonal effects should be conducted before SPC testing [70]. Data decomposition techniques, such as the Seasonal Trend decomposition procedure based on Loess (STL), are used to remove these effects [71,72].

Finally, data reformatting and data transformation are essential for machine learning-based methods. For instance, converting input data streams into tapped delay line formats facilitates certain models [61]. Transforming time-series data into hourly or minute time steps is a common preprocessing step for prediction classification, pattern identification, and clustering. These steps ensure that the data structure aligns with the requirements of specific analytical methods, enhancing model performance and reliability.

3.3.3. Performance Evaluation

After constructing a leak detection model, it is crucial to evaluate its performance and accuracy using a comprehensive test dataset. Figure 9 illustrates the Venn diagram summarizing the distribution of test data employed in various studies. The test data includes real leak events, simulated burst scenarios, or a combination of both. Simulated events are often designed to replicate real-life situations, such as the opening of fire hydrants [9], drain valves [8,76], or during hydrant flushing activities [62]. Moreover, hydraulic modeling tools, including the Water Network Tool for Resilience (WNTR) and EPANET, can generate artificial burst scenarios under controlled conditions.

Performance evaluation metrics are crucial for assessing model effectiveness. As shown in Figure 10, the most commonly used metrics (26.66%) include True Positive Rate (TPR), False Positive Rate (FPR), True Negative Rate (TNR), and False Negative Rate (FNR). TPR, also called recall or sensitivity, measures the proportion of actual leaks accurately detected, with a higher TPR indicating better performance. FPR measures the rate of false alarms, where a lower value is preferable for reliable detection. About 26.66% of reviewed studies rely solely on True Positives (TP) for validation, which, while effective at confirming actual leaks, does not address false alarms, limiting their overall reliability. A detailed summary of these metrics is provided in

Table 4. Performance evaluation metrics.

Performance Indicators	Acronym	Definition	Formula	Preferable Value
True Positive	TP	Number of instances correctly classified as leaks	$TP=\sum Positive prediction$	High
True Negative	TN	Number of instances correctly classified as non-leaks	$TN=\sum Negative prediction$	High
False Positives	FP	Number of instances wrongly classified as leaks	$FP=\sum False Positive prediction$	Low
False Negatives	FN	Number of instances wrongly classified as non-leaks	$FN=\sum False Negative prediction$	Low
True Positive Rate	TPR	Rate by which the real-life leaks are correctly predicted leaks	TPR $={\displaystyle \frac{TP}{TP+FN}}$	High
Positive Predictive Rate	FPR
Detection Probability	DP
Recall Rate	RR
False Positive Rate	FPR	Rate by which non-leaks are predicted as leaks	FPR $={\displaystyle \frac{FP}{TN+FP}}$	Low
Rate of False Alarms	RF
True Negative Rate	TNR	Rate by which non-leaks are correctly predicted as non-leaks	TNR $={\displaystyle \frac{TN}{TN+FP}}$	High
Specificity	SPC
False Negative Rate	FNR	Rate by which leaks are wrongly predicted as non-leaks	FNR $={\displaystyle \frac{FN}{TP+FN}}$	Low
Loss Alarm Rate	LAR
Average Detection Time	ADT	The average value of time taken for detection	ADT $={\displaystyle \frac{{\sum }_{i=1}^n{T}_i}{n}}$	Low
Accuracy	ACC	The ratio of correctly classified events to the total number of events	TPR $={\displaystyle \frac{TP+TN}{TP+FN+TF+FP}}$	High
Positive Predictive Value	PPV	Precision or the correct answer rate of the model	PPV $={\displaystyle \frac{TP}{TP+FP}}$	High
Mean Absolute Percentage Error	MAPE	Average percentage difference between predicted values and actual values	MAPE $={\displaystyle \frac{100}{n}}{\sum }_{i=1}^n\left|{\displaystyle \frac{{Actual}_i-{Predicted}_i}{{Actual}_i}}\right|$	Low

.

3.4. Leak Monitoring Platforms

Once data have been collected and processed, effective Human–Machine Interfaces (HMIs) are essential to support user interaction, real-time monitoring, and decision-making in SWM systems. These interfaces allow stakeholders and users to view processed data, receive alerts, and analyze network performance for timely intervention. For instance, Schultz et al. [67] investigated the use of a web portal that provided water consumption and leak alerts, finding that customers who used the portal responded more promptly to leaks than those who did not, highlighting the importance of accessible interfaces in water efficiency. However, only 35% of the reviewed studies focused on developing a system combining databases, alerts, and user interfaces.

Based on their functional characteristics, the leak monitoring platforms reviewed in the literature can be broadly categorized into three types: real-time monitoring platforms, GIS-integrated platforms, and digital twin-enabled platforms. Real-time monitoring platforms are designed to detect leaks as they occur and provide immediate alerts to users. Event detection systems like the Emergency Response System (ERS) [60] and the Dynamic Linear Models (DLM) tool [56] represent this category, enhancing emergency capabilities by identifying pipe bursts, illegal connections, sensor faults, and anomalies. These platforms often calculate event probabilities and trigger alarms when thresholds are exceeded. Basic platforms combining databases with interfaces and alerting functionalities also fall under this category. For example, Mounce et al. [25] proposed a platform integrating MS Access with online GUI and email alerts, enabling seamless coordination with customer service and maintenance workflows.

From a deployment perspective, real-time monitoring systems can be categorized into edge-based, cloud-based, and hybrid. Edge and IoT-based platforms, such as the Node-Red setup, integrating flow sensors and cameras with a Raspberry Pi 4 [26] process data locally for real-time alerting without depending on a central server. Cloud-based platforms offer scalability and advanced analytics. For instance, Shayp’s leak monitoring system [21] and the Linux-based SBC system using AWS and Rilheva [30] focus on indoor water monitoring with cloud-based alerting and archiving. Additionally, Boudville et al. [44] used the Blynk app with a Google Sheet backend for real-time dashboards and email notifications. Hybrid architectures, such as LoRaSURFING, combine IoT and cloud components to simulate and monitor distributed water networks in real-time, bridging local responsiveness with centralized processing capabilities [20]. Similarly, Fikejz et al. [27] developed mobile and web apps with MySQL backends compatible with Java and iOS, enhancing accessibility.

While all platforms support real-time monitoring, GIS-integrated platforms enhance this functionality by embedding spatial data, enabling precise geolocation of leaks. For instance, Cantos et al. [34] used GIS to improve leak visualization and situational awareness, demonstrating the value of spatial context in water infrastructure monitoring. These systems support better planning, response coordination, and asset management by visualizing leak locations directly on maps.

Going further, digital twin-enabled platforms serve as dynamic digital replicas of the physical system. These platforms enable bi-directional communication with physical assets, allowing for monitoring, simulation, and control. For instance, Laucelli et al. [80] introduced digital twin interfaces, such as DigitalWaterVirtualDMA_Design and DigitalWaterSegment_Viewer, for pipeline rehabilitation. Similarly, Wu et al. [66] developed a system integrating real-time sensor data with simulation tools, facilitating physics-based and data-driven analyses to detect and localize anomalies in water grids. While GIS and real-time systems serve as digital shadows, digital twins extend functionality by enabling predictive analytics and operational optimization. The functional characteristics of the leak management platforms are summarized in

Table 5. Functional characteristics of leak monitoring platforms.

Platform Type	Studies	Visualization	Control	Spatial Context	Simulation
Real-time monitoring	[20,21,25,26,27,30,44,56,60]	Dashboards, web/mobile GUIs	Not supported	Not supported	Not supported
GIS-based	[34]	Spatial map-based interfaces	Not supported	Leak geolocation	Not supported
Digital Twin enabled	[66,80]	Advanced dashboards + virtual models	Supports two-way control	Leak geolocation	Not supported

.

Prompt alerts are essential for addressing anomalies such as leaks, and various studies have proposed different alarm mechanisms based on system architecture and application needs. Mounce and Boxall [32] implemented an ANN/FIS-based Decision Support System (DSS) that generated email alerts for leak detection, demonstrating the use of intelligent algorithms for decision-making. Affifi et al. [22] utilized MQTT IoT protocols to deliver alerts via a web server, emphasizing lightweight communication for real-time responsiveness. Similarly, Choudhary et al. [35] adopted an SCADA-based system using Factory Talk View Studio and sensor integration (pressure, temperature), by providing email notifications. In contrast, Kane et al. [45] developed a real-time client for server-based alert triggering, while Bakker et al. [57] focused on alarm systems tailored for online applications. Mounce et al. [77] presented the semi-automated AURA system, which provides early leak warnings through automated monitoring and human oversight.

These examples highlight the wide range of alert mechanisms adopted in existing studies. In future research, selecting an appropriate alert mechanism should consider several key factors, including the network architecture (e.g., centralized vs. edge processing), required response times, and the specific needs of different end users. Alerts may take the form of system notifications on centralized dashboards, email alerts, or SMS messages. To ensure effective implementation, future studies should be guided by the following principles: real-time capability, enabling minimal latency between leak detection and alert generation; user specificity, ensuring alerts are tailored to the roles and responsibilities of different stakeholders; and integration potential, allowing for seamless connection with existing systems such as SCADA, GIS, or digital twin platforms. Selecting the right mechanism based on these criteria will be critical to enhancing the responsiveness and reliability of automated leak management systems.

4. Discussion

This section explores research questions by examining advancements in automatic leak management systems. The analysis underscores the critical role of selecting appropriate data collection and analytics techniques tailored to the specific characteristics of the Water Distribution System (WDS) under study. These foundational stages significantly influence the accuracy and reliability of the leak detection models. Moreover, all subsequent steps in developing the leak management framework must align with the chosen data analysis methodology. The findings also emphasize the potential to customize data alerts and warnings to meet specific user requirements, enhancing the system’s practical applicability.

4.1. RQ1 What Are the Existing Methods of Leak Management Using Water Flow Data?

Research question 1 aims to analyze the existing methods of leak management using flow data. The systematic literature review revealed a wide range of methods for leak management that utilize water flow data. These methods can be broadly categorized into leak detection, leak assessment and size estimation, leak localization, and leak prediction. While most studies focus on leak detection, some combine detection with other aspects, such as localization, size estimation, or assessment. Each method applies different data analysis techniques tailored to specific leak management objectives.

Table 6. Advantages and challenges of leak detection techniques.

Approach	Advantage	Challenge
Water Balance Approach	Easy to implement	It gives only a crude estimate of leaks Underestimate apparent loss
Minimum Night Flow approach	Provides a reasonable estimate of leaks	It does not provide a real-time leak warning
	Less computational power
	It does not need historical data.
Classification approach	Performs well when trained with simulated data created under different leak scenarios	Building a non-linear model that completely describes the hydraulic condition is impossible
		Requires historical data with balanced leak and non-leak scenarios
		The lack of labels in hydraulic data makes training difficult.
Prediction-classification approach	It does not require leak data for training	Historical data are highly unstable, and they require data preprocessing
		It needs a massive amount of historical data.
		Does not perform well in non-stationary conditions
Statistical approach	It does not need historical data	Results can be highly unstable
	It does not require accurate modelling of the WDS	Does not perform well in non-stationary conditions
	Less computation costs
Clustering techniques	It does not require prior knowledge of leaks for practical application	Needs historical data
	It does not require historical leak data	High computational and data storage efficiency
		Chance of high FPR

summarizes the strengths and limitations of different leak detection techniques reported in the literature. It is important to note that these techniques were applied in diverse system configurations and varying assumptions, input data types, and evaluation criteria. Hence, the table is not a direct comparison but an illustrative summary to help understand the context in which each method can be used.

Leak assessment methods, such as the water balance and Minimum Night Flow (MNF) approaches, provide insights into overall water loss within a WDS. The water balance method primarily overviews actual and apparent losses but tends to underestimate apparent losses [84]. As such, it often serves as a preliminary step, requiring integration with other methods to provide detailed leak volume estimates [24].

The MNF approach offers a simpler and computationally efficient alternative. Analyzing nighttime flow data can identify continuous leaks and even localize them when combined with data from multiple meters. However, its focus on nighttime data limits its ability to detect sudden bursts or leaks during high-demand daytime hours. Despite these limitations, MNF is widely applied due to its low data requirements and ability to combine leak detection with assessment when smart meter data are available.

Leak detection is the most widely researched area, leveraging both supervised and unsupervised approaches. Supervised techniques, such as classification and prediction-classification methods, require the hydraulic system to be initially modeled. However, the inherent complexity and uncertainty of WDS make accurate modeling challenging. Supervised classification also relies on historical data with a balanced representation of leak and non-leak scenarios, which may not always be available. To overcome this, some studies use simulated data from hydraulic simulation software to train models. However, manual labeling of leak scenarios remains a limitation of these methods.

Prediction-classification methods address challenges by requiring only normal hydraulic data to create prediction models. Hence, this method is widely used in leak detection. These models, however, depend heavily on substantial amounts of historical data, which may be noisy or incomplete, potentially reducing prediction accuracy. Data filtering techniques are, therefore, crucial in these approaches. Additionally, newly constructed pipe networks often lack sufficient historical data, and prediction models may struggle to adapt to evolving hydraulic conditions or non-stationary factors such as changes in water demand [8].

In contrast, unsupervised methods like SPC and clustering do not require historical leak data. This makes them advantageous for real-world applications where leak conditions are unpredictable. However, these methods face challenges such as high False Positive Rates (FPR) caused by unexpected user activity. Modifications, such as time delay analysis or clustering enhancements [8], have been introduced to address these issues, significantly reducing FPR. Studies suggest incorporating hydraulic operation changes into SPC models can improve detection accuracy. Pressure data, being more sensitive to system disturbances, have proven more effective than flow data for SPC-based detection [69].

Leak localization predominantly relies on pressure data, which are more sensitive to system disturbances than flow data. Pressure reading can be used to calculate nodal pressures, but this requires hydraulic modeling, which adds complexity. Studies have shown that integrating real-time pressure data with other methods improves leak localization accuracy, making it an essential complement to flow-based detection approaches.

Overall, the reviewed methods strongly rely on water flow data, with varying degrees of integration with pressure data, modeling, and historical trends. Supervised methods are well suited for systems with abundant historical data but require careful calibration to avoid issues with imbalance and noise. Unsupervised methods excel in detecting unexpected leaks without prior knowledge, but must address challenges like FPR. Meanwhile, leak assessment methods provide a foundation for understanding water loss but often lack real-time applicability. Finally, pressure data emerge as a critical component for effective leak localization.

These findings highlight the importance of selecting techniques that align with the specific characteristics of the WDS, the available data, and the goals of the leak management strategy. While flow data remain central, integrating additional variables, such as pressure and operational changes, enhances the effectiveness of these methods.

4.2. RQ2 How Can the Methods Addressed in RQ1 Be Applied to Automated Real-Time Leak Management?

Research question 2 addresses how the methods mentioned in RQ1 compare with each other for application in automated real-time leak management. An effective automatic leak detection system must possess some essential features: real-time automatic leak detection, advanced analytical capability, system adaptability, and high accuracy. The methods identified in RQ1 vary significantly in their suitability for automated real-time leak management, with some being more conducive to real-time detection and others requiring additional systems or hybrid approaches. IoT and AMI facilitate real-time data acquisition and transmission by integrating smart meters, flow and pressure sensors, and communication modules. Leak detection methods capable of processing such data in real-time, particularly AI-enhanced data-driven techniques, are more appropriate for dynamic and automated systems.

AI-based approaches, like supervised approaches (classification and prediction classification) and unsupervised clustering techniques, can also adapt to the non-stationary hydraulic conditions of water distribution systems based on real-time data. These methods reduce the false alarms and enhance the accuracy by learning from historical patterns of the WDS, adapting to new leak scenarios and demand changes. However, they require periodic retraining and access to high-quality, denoised datasets to maintain this adaptability. Thus, their effectiveness depends on the detection systems’ data quality, infrastructure, and integration.

Despite these challenges, AI-based techniques offer significant advantages for automated real-time leak detection, particularly in mature systems. These techniques can process large volumes of real-time flow data with minimal human intervention, making them suitable for continuous monitoring. On the other hand, though simpler to implement, statistical approaches tend to exhibit higher False Positive Rates (FPR). As a result, they are generally more effective when combined with other data-driven methods to enhance overall accuracy and reliability.

One of the key methods discussed in RQ1 is the Minimum Night Flow (MNF) approach, which is effective for leak assessment but unsuitable for real-time leak detection. MNF relies on analyzing nighttime flow data, which makes it well suited for identifying continuous, long-lasting leaks but ineffective for detecting sudden bursts or leaks during peak demand hours. As a result, while MNF can be part of a comprehensive leak management strategy, it must be combined with other methods to facilitate real-time leak detection. For example, coupling MNF with water balance methods or more advanced leak detection techniques can allow for more precise leak size estimation.

Hybrid methods combine flow and pressure sensor data and are emerging as one of the most promising solutions for automated real-time leak management. While flow sensors can detect leaks and assess their magnitude, pressure sensors are essential for localizing leaks within the network. Hybrid systems that integrate data from these sensors can provide near real-time insights into the location and severity of leaks, making them highly effective for real-time leak management. However, these systems require significant computational resources to process the data in real-time, which can be a barrier in resource-constrained environments. Thus, while hybrid systems have high potential, their deployment depends on the availability of sufficient computational infrastructure.

In conclusion, methods that combine data-driven techniques, smart water meters, and pressure sensors offer the best prospects for automated real-time leak management. However, the effectiveness of these methods depends on factors such as data quality, computational resources, and the integration of real-time alert systems. While challenges remain, particularly in resource-limited environments, advancements in technology and infrastructure will continue to enhance the ability to detect and manage leaks in real-time.

4.3. RQ3 What Is the Strength of the Evidence in Support of the Different Methods?

Research question 3 investigates the quality of evidence to support different studies. The quality of evidence varied across studies, with many relying on simulated events or engineered test events rather than real-world data. Only 30% of the reviewed studies were based on real-life leak events. Simulation-based studies provide controlled environments for testing leak management algorithms, but their generalizability to operational water distribution systems remains uncertain. Experimental studies with real-world leak events were more robust, offering practical insights, though such studies were limited in number. For machine learning methods, issues such as dataset size, diversity, and transparency in algorithm validation were observed.

A major challenge in comparing leak detection studies lies in the inconsistent application of evaluation metrics. Many studies solely report true positives, disregarding the importance of assessing false alarms. Furthermore, only a small fraction (7%) of studies consider average detection time a crucial factor for the effectiveness of real-time systems. Among studies that proposed leak detection systems, 37.5% did not utilize any evaluation metrics, while 16% limited their assessment to true positives. Moreover, only 12% explicitly addressed the critical aspect of leak detection time. These findings highlight the need for standardized evaluation procedures to ensure the reliability and comparability of leak detection research.

4.4. RQ4 What Implications Will These Findings Have with Respect to a Real-Time Automatic Leak Management System?

Achieving an automatic leak detection system requires seamless integration of leak detection techniques and flow sensor readings with emerging technologies, such as the Internet of Things (IoT), Artificial Intelligence (AI), and advanced leak monitoring platforms.

An important consideration for real-time leak management is the data collection infrastructure. While AMI and IoT-enabled smart water meters are key to real-time leak detection, not all data collection systems are designed for real-time monitoring. AMR systems, for example, provide high-frequency data but do not support real-time communication, thus limiting their use in automated leak management. While valuable for storing and analyzing historical data, flow meters with data loggers are generally incapable of providing real-time monitoring unless integrated with communication technologies that enable real-time transmission. Therefore, enabling continuous and immediate data flow is essential for an effective real-time leak detection system.

As discussed in Section 4.2, AI-enhanced leak detection techniques are particularly suitable for automatic leak management. The findings also emphasize the critical role of integrating multiple methodologies to develop an effective, automated leak management system. For instance, combining top-down approaches for system-wide monitoring with bottom-up methods for localized detection can significantly improve accuracy and reliability. However, to ensure the performance of such systems, it is essential to maintain high data quality through preprocessing techniques like denoising and reconstruction. Moreover, deploying such systems requires careful attention to infrastructure limitations, especially in resource-constrained areas, where the scalability of these technologies may be challenged.

Once an effective burst detection model is established, it is crucial to implement an automatic system for real-time leak alerts. While many of the discussed methods can detect leaks, the system must also be capable of issuing real-time alerts to relevant stakeholders. Integrating automated data analysis with alert mechanisms ensures that leak detection systems can trigger immediate actions, such as initiating repairs or alerting customers, without manual intervention. This is critical for minimizing water loss, reducing damage, and enhancing the overall efficiency of leak management efforts.

Despite the growing body of literature on this topic, few sources have comprehensively addressed the development of fully integrated systems for automatic leak management. For instance, Mounce et al. [25] developed an online AI-based burst and abnormal flow detection system that effectively handles data from multiple sites using MATLAB. Similarly, Romano et al. [60] introduced an Event Recognition System (ERS) to detect various events, including pipe bursts and sensor failures, by processing time-series data and triggering alarms when event probabilities exceed defined thresholds.

Moreover, Mounce et al. [77] introduced the AURA Alert system, which enhances the alert generation process by selecting high-quality training data and validating it against predefined thresholds. Hosted on the cloud, this system enables secure, real-time communication of alerts. These contributions illustrate how advanced analytics and real-time communication can significantly support utilities in leak detection. In addition, such integrated systems lay the groundwork for digital twins of water utilities, enhancing their capabilities to simulate, monitor, and manage conditions in real-time.

5. Conclusions

In conclusion, this systematic literature review delves into the advancements and challenges of developing automatic leak management systems using flow data. A total of 71 publications were reviewed, covering various methodologies for leak detection, localization, prediction, and assessment in WDS. The review identified three major components for developing an automatic leak detection system: data acquisition and management, the leak management model, and the leak-monitoring platforms. While a significant portion of the research focused on developing algorithms for leak detection, fewer studies addressed the broader scope of automated leak management systems, with most of the work concentrating on improving leak detection.

Automatic data collection is the core pillar of such systems. The review identified a growing trend in leveraging Internet of Things (IoT) technologies for real-time data acquisition. This includes integrating smart water meters with IoT and Advanced Metering Infrastructure (AMI) systems. Some sensors are coupled with edge processing devices, which can perform data processing near the source rather than transmitting the data to a centralized cloud. Once the data are collected, the data have to be sent to a central database for data processing.

The reviewed leak management techniques include leak detection, localization, and assessment. The review reveals a growing trend toward data-driven approaches, particularly AI-based models involving supervised and unsupervised approaches for automating leak detection and management in real-time. These models, often combining flow and pressure data, demonstrate significant potential for improving the accuracy and responsiveness of leak management systems. However, challenges persist, including the dependency on high-quality data, the need for substantial computational resources, periodic retraining, and the need for robust infrastructure to support real-time deployment.

Furthermore, the review emphasizes the importance of integrating multiple methods to build a comprehensive and robust leak management system. Combining traditional approaches, like water balance and MNF, with modern AI-driven techniques can improve system performance. Integrating smart metering, IoT platforms, and machine learning enables continuous system monitoring and facilitates real-time insights. Nevertheless, data quality remains a critical challenge, especially in newly developed or resource-constrained areas where historical data may be sparse or unreliable. The review suggests that future research should address these challenges, improving model accuracy, reducing false alarms, and enhancing real-time scalability.

Equally important is developing leak monitoring platforms equipped with real-time visualization and alert mechanisms. These platforms can be broadly classified as real-time leak monitoring, GIS-based, and Digital Twin-based platforms. GIS-based platforms facilitate spatial localization of leaks, while Digital Twins offer the most advanced functionality by enabling bi-directional communication between the virtual and physical water networks. Although these platforms have significant potential to reduce detection and response times, further research is needed to develop scalable, user-friendly solutions.

Integrating smart metering technologies, IoT, and advanced monitoring platforms presents a promising path forward for enhancing the efficiency and sustainability of global water distribution systems. In summary, while substantial progress has been made in developing automated leak management systems, continued technological and infrastructural advancements are essential to unlock their potential in real-world applications fully.