Author(s): Nelson Carrico; Joao Salgado Silva; Joao Caetano; Didia Covas
Linked Author(s): Dídia Isabel Cameira Covas
Keywords: Anomaly detection; Autoencoder; DMA inflow; Matrix profile; Streaming
Abstract: Real-time anomaly detection in District Metered Areas (DMAs) is critical for minimising non-revenue water and enabling rapid operational response. This study evaluates two unsupervised machine learning methods – Autoencoder and Matrix Profile – implemented within a streaming framework using high-frequency telemetry data (5-minute intervals) from a DMA in Vilamoura, Portugal. The pilot area is characterised by pronounced seasonal variability, where water demand increases by a factor of five during summer months due to tourism. This severe concept drift, combined with the noise inherent in high-resolution sampling, poses a challenge: rapid baseline shifts and stochastic consumption spikes can easily mimic or mask true anomalies. The comparative analysis assesses the sensitivity, stability, and computational suitability of each method under these highly dynamic hydraulic conditions. The findings are expected to support the development of adaptive, resource-efficient monitoring systems capable of handling the load fluctuations typical of tourist-centric water networks.
Year: 2026