SUNWater

Abstract The water supply network (WSN) is subjected to leaks that compromise its service to the communities, which, however, is challenging to identify with conventional approaches before the consequences surface. This study developed Machine Learning (ML) models to detect leaks in the WDN. Water pressure data under leaking versus non-leaking conditions were generated with holistic WSN simulation code EPANET considering factors such as the fluctuating user demands, data noise, and the extent of leaks, etc. The results indicate that Artificial Neural Network (ANN), a supervised ML model, can accurately classify leaking versus non-leaking conditions; it, however, requires balanced dataset under both leaking and non-leaking conditions, which is difficult for a real WSN that mostly operate under normal service condition. Autoencoder neural network (AE), an unsupervised ML model, is further developed to detect leak with unbalanced data. The results show AE ML model achieved high accuracy when leaks occur in pipes inside the sensor monitoring area, while the accuracy is compromised otherwise. This observation will provide guidelines to deploy monitoring sensors to cover the desired monitoring area. A novel strategy is proposed based on multiple independent detection attempts to further increase the reliability of leak detection by the AE and is found to significantly reduce the probability of false alarm. The trained AE model and leak detection strategy is further tested on a testbed WSN and achieved promising results. The ML model and leak detection strategy can be readily deployed for in-service WSNs using data obtained with internet-of-things (IoTs) technologies such as smart meters.

Abstract Water distribution networks (WDNs) are critical infrastructure for communities. The dramatic expansion of the WDNs associated with urbanization makes them more vulnerable to high-consequence hazards such as earthquakes, which requires strategies to ensure their resilience. The resilience of a WDN is related to its ability to recover its service after disastrous events. Sound decisions on the repair sequence play a crucial role to ensure a resilient WDN recovery. This paper introduces the development of a graph convolutional neural network-integrated deep reinforcement learning (GCN-DRL) model to support optimal repair decisions to improve WDN resilience after earthquakes. A WDN resilience evaluation framework is first developed, which integrates the dynamic evolution of WDN performance indicators during the post-earthquake recovery process. The WDN performance indicator considers the relative importance of the service nodes and the extent of post-earthquake water needs that are satisfied. In this GCN-DRL model framework, the GCN encodes the information of the WDN. The topology and performance of service nodes (i.e., the degree of water that needs satisfaction) are inputs to the GCN; the outputs of GCN are the reward values (Q-values) corresponding to each repair action, which are fed into the DRL process to select the optimal repair sequence from a large action space to achieve highest system resilience. The GCN-DRL model is demonstrated on a testbed WDN subjected to three earthquake damage scenarios. The performance of the repair decisions by the GCN-DRL model is compared with those by four conventional decision methods. The results show that the recovery sequence by the GCN-DRL model achieved the highest system resilience index values and the fastest recovery of system performance. Besides, by using transfer learning based on a pre-trained model, the GCN-DRL model achieved high computational efficiency in determining the optimal repair sequences under new damage scenarios. This novel GCN-DRL model features robustness and universality to support optimal repair decisions to ensure resilient WDN recovery from earthquake damages.

Leakages in the underground water distribution networks (WDNs) waste over 1 billion gallon of water annually in the US and cause significant socio-economic loss to our communities. However, detecting and localization leakage in a WDN remains a challenging technical problem despite of significant progresses in this domain. The progresses in machine learning (ML) provides new ways to identify the leakage by data-driven methods. However, in-service WDNs are short of labeled data under leaking conditions, which makes it infeasible to use common ML models. This study proposed a novel machine learning (ML)-based framework for WDN leak detection and localization. This new framework, named clustering-then-localization semi-supervised learning (CtL-SSL), uses the topological relationship of WDN and its leakage characteristics for WDN partition and sensors placement, and subsequently utilizes the monitoring data for leakage detection and leakage localization. The CtL-SSL framework is applied to two testbed WDNs and achieves 95% leakage detection accuracy and around 83% final leakage localization accuracy by use of unbalanced data with less than 10% leaking data. The developed CtL-SSL framework advances the leak detection strategy by alleviating the data requirements, guiding optimal sensor placement, and locating leakage via WDN leakage zone partition. It features excellent scalability, extensibility, and upgradeability for applications to various types of WDNs. It will provide valuable a tool in sustainable management of the WDNs.

Underground water pipes deteriorate under the influence of various physical, mechanical, environmental, and social factors. Reliable pipe failure prediction is essential for a proactive management strategy of the water supply network (WSN), which is challenging for the conventional physics-based model. This study applied data-driven machine learning (ML) models to predict water pipe failures by leveraging the historical maintenance data heritage of a large water supply network. A multi-source data-aggregation framework was firstly established to integrate various contributing factors to underground pipe deterioration. The framework defined criteria for the integration of various data sources including the historical pipe break dataset, soil type dataset, topographical dataset, census dataset, and climate dataset. Based on the data, five ML algorithms, including LightGBM, Artificial Neural Network, Logistic Regression, K-Nearest Neighbors, and Support Vector Classification are developed for pipe failure prediction. LightGBM was found to achieve the best performance. The relative importance of major contributing factors on the water pipe failures was analyzed. Interestingly, the socioeconomic factors of a community are found to affect the probability of pipe failures. This study indicates that data-driven analysis that integrates the Machine Learning (ML) techniques and the proposed data integration framework has the potential to support reliable decision-making in WSN management.

Predicting the time-dependent pipe failure rate of the water distribution networks (WDNs) is important for planning its renewal budget but also challenging due to the complex factors involved. The recent development of machine learning techniques provides a novel approach for accurate failure rate prediction based on historical data. However, the inherited randomness and uncertainty of water pipe failures and machine learning algorithms are often ignored in the training and prediction process. This article develops an uncertainty quantification framework for deep learning-based WDN system-wide failure rate prediction. The framework integrates a probabilistic long short-term memory (LSTM) model with a Monte Carlo method. The historical climate data and WDN pipe maintenance data over the past 35 years for Cuyahoga County, USA, are used to illustrate this deep-learning model-based uncertainty quantification framework in this study. A statistical time series regression model, ARIMAX, is used as a comparison benchmark. The results show that the LSTM model outperforms the ARIMAX model in prediction accuracy in most years by considering the uncertainties. Besides, the uncertainty range of the LSTM model prediction is 50% of that of the ARIMAX model. The results also identified the major contributing factors to the uncertainties of LSTM machine learning model prediction. The proposed uncertainty framework features excellent extensibility and can be adapted to quantify uncertainties with other types of machine learning models.

Abstract Efficient representation of complex infrastructure systems is crucial for system-level management tasks, such as edge prediction, component classification, and decision-making. However, the complex interactions between the infrastructure systems and their spatial environments increased the complexity of network representation learning. This study introduces a novel geometric-based multimodal deep learning model for spatially embedded network representation learning, namely the regional spatial graph convolutional network (RSGCN). The developed RSGCN model simultaneously learns from the node’s multimodal spatial features. To evaluate the network representation performance, the introduced RSGCN model is used to embed different infrastructure networks into latent spaces and then reconstruct the networks. A synthetic network dataset, a California Highway Network, and a New Jersey Power Network were used as testbeds. The performance of the developed model is compared with two other state-of-the-art geometric deep learning models, GraphSAGE and Spatial Graph Convolutional Network. The results demonstrate the importance of considering regional information and the effectiveness of using novel graph convolutional neural networks for a more accurate representation of complex infrastructure systems.