Urban flooding, intensified by climate change and urbanization, poses significant threats to public safety and economic growth, with annual costs exceeding 100 billion. This research addresses this pressing issue by introducing a transformative approach to urban flooding forecasting through advanced artificial intelligence. Central to our innovation is a segmented deep learning model that effectively captures the diverse characteristics of urban environments, facilitating more accurate flood predictions. Its fine-tuning capabilities enable the model to adapt to real-time data and dynamic urban conditions, resulting in significantly enhanced predictive performance. This advancement not only surpasses the accuracy of traditional forecasting methods but also offers substantial potential for real-time applications, supporting city planners and emergency responders in proactive flood management.

Urban water distribution networks (WDNs) require adaptive scheduling strategies to handle dynamic demand fluctuations. Scientific pump scheduling ensures reliable water supply while minimizing energy consumption and carbon emissions. Data-centric AI (DCAI) enables real-time intelligent scheduling of large-scale WDNs by learning human expertise from historical data, yet its effectiveness is often constrained by data quality and static model deployment. Most existing studies train models once for long-term use, neglecting continuous online updates and limiting practical adaptability; this constitutes a critical barrier for autonomous scheduling agents. To address this issue, this study proposes a data-model coevolution mechanism that transforms DCAI into an autonomous scheduling agent with two core capabilities: (1) a dynamic data curation system that ensures high-quality data streams through multi-criteria filtering, and (2) a data-coevolving model update mechanism with retraining or fine-tuning options for continuous online learning and self-evolution. Validated on a real network in Shanghai over 10-week cycles, the agent-based mechanism reduced MAE by up to 71.03% and improved water supply security and energy efficiency. Interpretability analysis confirmed its directional controllability during autonomous updates. Furthermore, when extended with an active learning variant, the proposed approach achieved greater than 25% MAE reduction across scenarios, addressing the inherent challenge faced by transfer and continual learning approaches in capturing minority operational patterns. By endowing DCAI with autonomous online learning capabilities, this work establishes a synergistic paradigm of intelligent scheduling in which the agent achieves automated adaptive updates through interaction with the WDN, thereby generating superior scheduling strategies.

Large language model (LLM)-based agents offer a promising paradigm for automating engineering workflows in water distribution network (WDN) simulations. However, existing frameworks generally assume that user instructions are directly executable, rely heavily on cloud-based inference, and often lack robust domain-specific semantic grounding, limiting their reliability and practical deployment. This study presents LightWDN-Agent, a lightweight multi-agent framework for natural language-driven WDN simulation that integrates quality-aware task validation, retrieval-augmented semantic grounding, and local–cloud collaborative inference. The framework first evaluates the executability of user queries and requests clarification for ambiguous or incomplete instructions before task execution. It then incorporates domain knowledge, tool descriptions, and historical successful cases through retrieval-augmented generation to constrain task planning and improve semantic consistency. Finally, routine reasoning is performed using lightweight local models, while cloud-based large language models are invoked only for complex reasoning and runtime error correction, enabling privacy-aware and resource-efficient deployment. The framework was evaluated using a benchmark comprising 39 representative WDN simulation tasks and 117 natural language queries with different quality levels. The proposed quality-gating strategy effectively prevented ambiguity propagation, improving execution reliability while reducing redundant reasoning. Threshold calibration identified an appropriate operating range that balances task acceptance and execution success. Approximately 69.3% of inference tokens were processed locally, with 38.5% of tasks completed entirely without external API calls. In addition, removing retrieval-augmented semantic grounding reduced the task completion rate from 100% to 46.15% and substantially increased token consumption. The results demonstrate that reliable natural language-driven WDN automation requires not only strong language models but also executability regulation, structured semantic grounding, and deployment-aware inference allocation, providing a practical framework for intelligent and privacy-preserving water system management.

Urban sewer systems increasingly require intelligent, low-cost monitoring strategies to maintain environmental safety and operational resilience. This study proposes a spatiotemporal graph learning framework for detection and localization of external inflows in urban sewer systems. The framework combines three inputs: (i) the inflow-affected water-level time series from one downstream sensor, (ii) simulated baseline water-level series for all upstream nodes generated using EPA SWMM, and (iii) a weighted graph representation of network topology. An LSTM-GCN regression module first estimates the event-level maximum water-level deviation at each node rather than the full deviation time series, improving robustness to unknown inflow hydrographs. A downstream GCN classifier then converts the inferred spatial response into branch-level probabilities for source localization. The method was evaluated in a synthetic 117-junction test network and a 39-junction real-world sewer subnetwork. In the synthetic single-source case, the true source branch was ranked within the top two candidates in over 90% of test events. In the real-world case, the Top1 branch localization reached 100%. Sensitivity analyses further showed that larger inflow magnitudes and hydraulically coherent branch partitioning improved localization reliability. These results demonstrate the potential of graph-based single-sensor diagnostics for low-cost sewer monitoring and targeted inspection planning.

This study develops an adaptive urban flood forecasting framework by integrating hydrodynamic modeling and deep learning. A coupled hydraulic model is first used to generate physically consistent training data that jointly represent subsurface network dynamics and surface inundation processes. Based on these data, the framework combines a Crossformer model for forecasting global nodal water levels, an autoencoder for compressing high-dimensional inundation fields, and a multilayer perceptron that maps subsurface hydraulic states to the latent features of surface flooding. To support operation under sparse monitoring conditions, inversion models are introduced to reconstruct global 1D and 2D flood states from limited sensor observations, and an online continual learning strategy is used to update the forecasting model during deployment. Results show that the proposed framework achieves high predictive skill for both subsurface and surface flooding, with NSE values above 0.97 for nodal water levels and CSI values above 0.93 for inundation extent, while the online updating strategy reduces RMSE by about 40% during extreme rainfall events. These results demonstrate that the proposed framework provides an effective and operationally practical route toward adaptive urban flood forecasting under coupled sewer-surface conditions and sparse monitoring constraints.

The Graphical Abstract shows the scanned inner wall surface of a pipe and the statistics of the protrusion height in the upper part, while in the lower part displays the random model of inner wall and the simulated hydraulic performance.The hydraulic performance of water supply pipes deteriorates significantly due to inner wall corrosion associated with aging. This study investigated the spatial characteristics of corrosion patterns and their hydraulic impacts based on nine aged pipes, by integrating three-dimensional (3D) scanning, random field modeling, and computational fluid dynamics (CFD) simulation. Inner wall morphology data were acquired via 3D scanning from nine segments of aged cast iron pipes retrieved from a city in northern China. Random field theory was applied to characterize the spatial variability of the scanned morphology. Subsequently, a random field model was developed to reconstruct inner wall profiles, which were then used in CFD simulations to evaluate the hydraulic performance of the corroded pipes. Results indicate that the inner wall morphology exhibits pronounced spatial anisotropy, with a significantly larger correlation length in the circumferential direction than in the axial direction. The heights of corrosion protrusions approximately follow a lognormal distribution. CFD simulations demonstrate that the hydraulic performance of the inner wall morphology reconstructed by the proposed random field model closely matches that of the originally scanned pipes. This confirms the model's effectiveness in statistically reproducing the hydraulic effects of inner wall roughness in aged pipes.HIGHLIGHTSA random inner-wall model for aged cast iron pipes was developed based on 3D laser scanning. Protrusion heights followed a lognormal distribution and exhibit pronounced spatial anisotropy. The random inner-wall model effectively reproduced hydraulic behavior.

Perception and prediction of water quality dynamics is crucial for optimizing water system management in wastewater treatment plants (WWTPs). Various machine and deep learning methods are employed in WWTPs for these purposes, yet they rely on extensive data in water quality for model training, leading to inefficient data usage and high costs. This study presents a novel method based on Koopman analysis and clustering for identifying representative water quality dynamic data in WWTP. The dynamic data considered in this study includes temperature, inflow, COD, TN, NH4+, NO3-, MLSS, and DO. These representative data can then be used to train different ML-based models, so as to reduce the data requirements without much decrease in their accuracy. This method is demonstrated for a A2O process of a WWTP in southern China. The ML models trained on the selected representative data are statistically comparable to those trained on the full time period dataset, while significantly reducing data requirements by 43.59%. This study offers an efficient and cost-effective framework for data management and model training in WWTPs and various water systems where data collection is resource-intensive. It also provides a new perspective that extensive data is not necessary for training ML models in water system when we acquire representative WWTP data capable of capturing all dynamical regimes.

Data-driven models have increasingly been used as useful tools for process simulation in urban wastewater treatment plants (WWTPs), but their lack of interpretability and limited generalization hinder their application in practical engineering scenarios. In this study, a hybrid neural network incorporating an attention-based encoder and a long short-term memory module was proposed for data-driven modeling of the anaerobic-anoxic-oxic (A2O) process in real WWTPs. Compared with several widely used baseline deep learning architectures for time-series prediction, the proposed method demonstrated improved accuracy and efficiency in predicting dissolved oxygen (DO) concentrations in the oxic zone. The model was further analyzed using a bidimensional interpretable deep learning framework, enabling global interpretability analysis across temporal and feature levels. This bidimensional perspective provided new insights for optimizing model architecture and feature selection. Three practical scenarios characterized by stepped water temperature distributions were defined, and local interpretability methods were employed to enhance understanding of the individual and interactive effects of aeration rate and water temperature on DO within each scenario. The results indicate that the interaction mechanisms among core variables affecting DO vary significantly across different seasonal conditions. Therefore, it is necessary to design scenario-specific aeration strategies to optimize real-time control operations in WWTPs. Additionally, the transfer learning method, incorporating dynamic system similarity, significantly enhanced the model's generalization capability. This suggests that transferring and retraining models across similar A2O systems may be an effective approach for improving model reuse and addressing the "cold start" challenge in newly constructed WWTPs.

Efficient and sustainable water supply is a critical element for the normal functioning of cities, with drinking water systems being vital to public welfare. In addition, rapid urbanization and climate change have led to increasingly complex water supply problems, which require new solutions from urban drinking water system authorities. As a result, many researchers focused on innovations within the urban water sector to ensure water security. Deep learning (DL) algorithms, as disruptive technologies, are increasingly being applied to planning and management issues in urban water supply systems. However, there was a lack of systematic and up-to-date review of the current state of DL algorithms applied in urban water systems from source to tap. Therefore, this review combined the applications of DL in urban water source management, drinking water treatment, water distribution networks, and building water supply systems based on literatures published in the Web of Science database from 2021 to 2025, covering the entire process of water supply monitoring, prediction, (anomaly) detection, control, optimization, and management. It is found that DL applications were significantly more prevalent in water resources and distribution networks than in water treatment and building water supply systems. Meanwhile, the integration of multi-agent systems and digital twins will be a hot research topic and major challenge for the future water industry. It is expected that this review will provide guidance for the development of intelligent urban drinking water systems, contributing to the sustainable development of water resource management and the digital transformation of the water industry.

To enhance the effectiveness and spatial rationality of routine water quality monitoring in water distribution networks (WDNs), this study proposes a multi-point optimization method that integrates information theory, multi-indicator collaboration, and structured deep clustering. The framework jointly considers two key water quality indicators, water age and residual chlorine. It further employs a Structured Deep Clustering Network (SDCN) to achieve high-consistency zoning. Candidate monitoring points are assessed by their multi-indicator information-entropy gain and their contribution to spatial coverage. A greedy algorithm is then used to construct an optimized sensor layout under spatial constraints. The method is applied to both the benchmark BWSN 1 network and a real-world water distribution system in MD Town, Jiangsu Province. Results show that the proposed approach outperforms existing layouts in terms of information acquisition, redundancy control, and spatial balance. It also demonstrates strong adaptability across monitoring scenarios and practical engineering feasibility. In particular, refined quality zoning and marginal-utility analysis lead to a deployment strategy that achieves an effective trade-off between monitoring cost and information representation. This provides a practical pathway for intelligent sensor deployment in complex WDNs.

Urban water environment pollution is a pressing global concern, particularly in developing countries where inadequate infrastructure contributes significantly to this challenge. This study builds upon these principles by enhancing drainage pipeline inspection technologies, aiming to streamline processes and reduce resource consumption. This research advances the integration of gyroscopes and accelerometers within a sextuple-axis sensor framework, streamlining a workflow wherein the inspection apparatus is introduced into the conduit and navigates with the fluid motion to aggregate data. The implementation of an attitude determination algorithm rooted in the extended Kalman filter underpins the processing of sensor-acquired data, yielding precise tridimensional attitude measures. Additionally, a refined peak-to-peak anomaly detection technique, based on an adaptive peak algorithm, analyzes the attitude measures to pinpoint deviations in the device's orientation. Empirical evaluations corroborate that the second-generation pipeline inspection device conceived in this study boasts augmented stability and transit efficacy. The integrated approach for attitude calculation and anomaly discernment coalesces data from gyroscopes and accelerometers, guaranteeing meticulous orientation angle computation and enhanced precision in anomaly detection. This accuracy is vital for the accurate replication of the detector's positioning within the pipeline infrastructure and for a comprehensive understanding of the operational state of drainage conduits.

Urban drainage systems (UDSs) are vital for managing stormwater and wastewater but face growing challenges due to urbanization, climate change and aging infrastructure. Real-time control (RTC) enhances UDSs' performance and circumvents the need for system upgrades through adaptive management and repurposing existing systems. Meanwhile, deep reinforcement learning (DRL) has emerged as a promising tool to improve decision-making, stability in the dynamic, nonlinear and dimensional environments. Recent studies demonstrate the potential of deep reinforcement learning control (DRLC) in flood mitigation, sewer overflow reduction, water quality management, and wastewater treatment optimization. While DRLC offers transformative opportunities for UDSs control optimization, its widespread adoption and real-world implementation requires long-term effort to address technical and practical gaps. This review systematically evaluates DRLC's progress in UDSs, summarizes the critical limitations, and proposes constructive insights, including data management, surrogate model design, benchmark frameworks construction, interpretability, safe control frameworks, and UDSs resilience enhancement to advance its future research.

Optimal pressure sensor placement is important for several purposes, including leakage detection, model calibration and state estimation. However, pressure data are required to serve multiple objectives simultaneously, and a versatile methodology to optimize sensor locations across varying network scales remains lacking. Existing approaches, mainly validated in small-scale networks, often face computational inefficiencies when applied to larger systems, limiting their practical utility. To address these gaps, this paper proposes a Projection-Sensitivity Based Multi-Optimal Sensor Placement (PSMOSP) framework. It is designed to identify robust and adaptable monitoring points for water distribution networks (WDNs) of diverse sizes. The framework uses a multi-objective genetic algorithm to resolve the Pareto frontier. There are two key objectives: more accurate calibration of model parameters and leakage localization. It improves computational efficiency through an analytical approach that constructs the pressure-roughness sensitivity matrix and parallel computation. Case studies on networks of different scales validate the framework, showing that it outperforms traditional methods in leak location and calibration accuracy, as well as spatial coverage. The results highlight how sensitivity-driven calibration and projection-based localization enhance the robustness of sensor placement in large-scale WDNs, offering a generalized methodology for online monitoring for tasks like pressure estimation, model calibration, and anomaly detection.

Pipeline cleaning is essential for maintaining water quality and safeguarding public health within water distribution networks. Although ice slurry pigging is widely adopted due to its effectiveness and environmental advantages, challenges remain in accurately quantifying cleaning efficacy and optimising maintenance intervals. This study develops a stress-driven wall material removal model calibrated using field-measured turbidity data, allowing for a comprehensive assessment of cleaning effectiveness based on post-cleaning shear resistance. The shear stress distribution induced by ice slurry along the pipeline is characterised using Computational Fluid Dynamics simulations combined with field-measured conductivity data. Validation through two real-world case studies confirms the model's capability to characterise wall material conditions, objectively evaluate cleaning performance, and determine optimal cleaning intervals. Results indicate that an optimal cleaning interval of approximately 24 months is achievable when integrating a post-cleaning shear resistance of 5.2 Pa with operational cost considerations. Proposed optimisation strategies, including increasing ice slurry volume fraction and cleaning velocity, significantly enhance the effectiveness of the cleaning process, extending the maintenance interval by up to 12 months compared to conventional practices. Overall, the findings establish a robust quantitative framework for systematic pipeline maintenance, contributing critical technical insights for augmenting water quality resilience in municipal water networks.

This study presents an innovative approach to real-time modeling of urban drainage networks, leveraging a highly accurate coupled one- and two-dimensional hydrodynamic model to generate a training dataset for node water levels. By employing global states inferred from monitoring points as model inputs, this study overcomes the limitations imposed by the scarcity of monitoring data and the challenge of capturing all node levels. The Crossformer algorithm, which simultaneously accounts for correlation at both temporal and feature scales, is applied to enhance the precision of simultaneous water level predictions across the network. Comparative analysis of different prediction patterns reveals that extending predictions based on a high-accuracy infrastructure offers more benefits than direct modifications to the algorithm's structure. In addition, this paper pioneered the application of online continuous learning concepts to update the prediction model in real time, achieving a balanced integration of measured and simulated data. Consequently, this paper establishes a complete monitoring-predicting-updating real-time simulation system for urban drainage networks.

The physics-informed neural network (PINN) method has been applied to solve water hammer equations in pipeline systems due to its ability of seamlessly integrate measurement data with conservation laws, offering advantages over traditional numerical method. However, existing PINN approaches require multiple neural networks to construct composite models for complex water distribution systems (WDS). This situation treats nodal information as boundary condition or labeled data during training, leading to a weaker robustness and a high demand for data. To address these issues, a hybrid water hammer model based on eXtended Physics-Informed Neural Networks for WDS (WDS-XPINN) is developed in this study. Unlike the standard PINN, WDS-XPINN incorporates the nodal mechanistic model directly into the loss function, enabling to synchronously train a unified neural network jointly through sparse augmented measurement data for pipeline system. Additionally, an adaptive weights method is introduced to improve model robustness by balancing the contributions of flowrates and pressures. The proposed WDS-XPINN is evaluated in two case studies: a series pipeline system with different operational events and noise perturbation, as well as a topological structure with looped and branched pipe. According to the simulation results and uncertainty analysis, the WDS-XPINN model demonstrates its excellent capacity of modeling fluid transient accurately in pipeline system, even without exact operational conditions or true pipe parameters.

Deep reinforcement learning (DRL) has been applied to real-time control (RTC) of urban drainage systems (UDSs), with impressive performance and efficiency in reducing urban flooding and combined sewer overflows (CSO). However, for complex UDSs, learning from scratch is time-consuming and difficult to converge. In this study we construct a knowledge-data fusion DRL training framework to integrate different types of UDS engineering experience and knowledge into DRL. We first convert knowledge of various types, via simulation, into knowledge-carrying data, which are subsequently used to pre-train a DRL agent using a newly designed offline supervised learning method. The pre-trained model is then fine-tuned using the well-established reinforcement learning procedure, achieving flooding and CSO mitigation comparable to that of fully-trained DRL but with nearly 90 % reduction in the training time, when applied to a benchmark UDS model. The quality of the knowledge used in pre-training influences the DRL performance, making its quality control critical to ensure RTC performance. Meanwhile, integrating diverse types of knowledge can effectively enhance the performance of the framework.

In the field of water supply management, multi-steps water demand forecasting plays a crucial role. While there have been many studies related to multi-steps water demand forecasting based on deep learning, little attention has been paid to the interpretability of forecasting models. Aiming to improve both the forecasting accuracy and interpretability of the model, a novel urban water demand forecasting neural network (UWDFNet) was presented in this paper. Compared with traditional deep learning models, it innovatively considered domain-specific prior knowledge from water supply management and incorporated the correlation relationship between different input variables into the design of the neural network structure, and verified the consistency between the knowledge learned by the model and prior knowledge through interpretability analysis. Additionally, a systematic performance evaluation was conducted and proved that UWDFNet possesses better accuracy and stability compared to other baseline models(e.g., gated recurrent unit network (GRUN), GRUN with a corrected Network (GRUN+CORRNet), GRUN+PID, GRUN+Kmeans).