Anomaly Detection and Classification in Water Distribution Networks Integrated with Hourly Nodal Water Demand Forecasting Models and Feature Extraction Technique

Abstract

Effective detection and classification of abnormalities, such as pipe bursts, leakage, illegal water use, and sensor failures, are critical for assisting water utilities in decision making, rapid response, and minimizing damage and disruption. This work presents a new flow data-based anomaly detection and classification method in water distribution networks. The method first establishes hourly nodal water demand forecasting models, then uses a unique integration of feature extraction technique of flow curve and convolutional neural network method to enable anomaly detection and classification from continually updated time window flow data. Verification progress from real and synthetic data of the case network shows that the proposed method can identify four common types of abnormal patterns in a fast and reliable manner with high recognition accuracy. The established models have self-learning capabilities, can process flow data in real time, and do not require hydraulic models to assist in analysis, which can be promising for wide practical applications in smart management of water distribution systems.