Author(s): Alireza Mahvelati Shamsabadi; Daeseung Kyung
Linked Author(s):
Keywords: Anomaly detection; Climate extremes; Cross-sectional dependency; Wastewater treatment
Abstract: Climate extremes alter how wastewater treatment plants behave. However, most anomaly detection tools still analyse each plant or variable separately. We quantified the cross-sectional structure which is neglected when the plants are treated as independent. Using unsupervised learning, and constructing multivariate features, we applied two dependency-aware detectors: (i) a cross-sectional LSTM autoencoder that learns normal 14-day joint trajectories of both plants, and (ii) a Gaussian copula model fitted to the simultaneous dependency structure of flows, effluent and climate covariates. Models were calibrated in an unsupervised way using the 99th percentile of training scores. Anomalies were almost absent on climatologically normal days, but the frequency increased during the hottest and wettest 10% of days, showing that cross-sectional, climate-conditioned anomaly detection can provide early warning of system-level stress even when individual effluent limits seem to be met.
Year: 2026