Data-driven methods for online leakage detection in water distribution networks

Abstract

Leakage detection is one of the important aspects of water distribution management. Water utilities are actively exploring alternative approaches to promptly and accurately detect leaks, aiming to mitigate water losses and minimise associated environmental and economic consequences. However, existing burst detection methodologies encounter challenges such as high false alarm rates and inadequate detection accuracy. Moreover, they often overlook diverse leakage types, such as gradual leakage events which pose a greater challenge due to their slow, incremental growth, yet their undetected presence leads to significant water loss. This thesis addresses these challenges by presenting research on the development and testing of three novel data analysis methodologies for automated real-time detection of both gradual leakages and burst events based on flow monitoring data in water distribution networks. Firstly, a gradual leakage detection method employing a multi-step forecasting strategy based on multilayer perceptron (MLP) is proposed. By incorporating long-term prediction results, this method captures trend variations caused by gradual leakages and subsequently, a smoothing technique highlights the presence of gradual leakage events. Application of this method to both synthetic and real datasets demonstrate its effectiveness. Secondly, a burst detection method based on local residual discrepancy with multiple deep learning models (MLP, long short-term memory, and sequence to sequence) is proposed. A novel framework is devised to generate an ensemble prediction result, enhancing the detection performance compared to traditional single-step prediction methods. The proposed approach exhibits robustness and superiority when tested on synthetic burst events within both synthetic and real datasets. Lastly, a lightweight statistical method based on empirical weighted moving average that is capable of detecting both gradual leaks and bursts is proposed. The method comprises three subsystems: pre-processing, online detection, and automated parameter updating. Its user-friendly yet powerful nature makes it advantageous. Application of this method to a full-year dataset in an online manner further underscores its efficacy. Overall, these methodologies represent significant contributions to the field, offering improved capabilities for timely and accurate detection of both bursts and gradual leakages in water distribution networks. By addressing existing challenges with innovative solutions, this research contributes substantially to the knowledge base, paving the way for more effective water management strategies.