Unsupervised Anomaly Detection in Urban Water Networks U sing a Hierarchical Deep Learning Model

Efficient anomaly detection in urban water distribution networks is crucial for sustainable resource management. Identifying abnormal patterns in time-series data from sensors is a common task within industrial settings and automated methods can enhance efficiency and reduce costs. However, supervised learning is often impractical due to the scarcity of labeled anomalous data in settings of practical interest. Consequently, unsupervised techniques are typically used, identifying anomalies as deviations from normal behavior. This paper introduces a framework that combines Long Short-Term Memory (LSTM) and Transformer encoder-decoder networks in a hierarchical structure to predict future water usage and detect anomalies in an unsupervised manner. Our method optimizes the use of available data within practical settings, where partial information might not be available at some points, ensuring high inference efficiency and enabling near real-time anomaly detection. It is adaptable to various data configurations, making it suitable for different environments. Our approach, applied to data from the metropolitan area of Barcelona, highlights its potential to refine early anomaly detection, prevent resource loss, and promote sustainable consumption.