Transformer Based Adaptive Drift Aware Anomaly Detection for Energy Consumption Time Series Using Forecasted Normal Behaviour

Abstract

This dissertation studies daily anomaly detection in 15 minute energy consumption time series under contextual variability and concept drift. The goal is to decide, at the end of each day, whether the observed consumption profile is still explainable by the valid recent behaviour of the same site, or whether it deviates in a way that should be inspected by an operator. The proposed solution develops a causal rolling protocol. For each complete day, the system selects recent historical days from the current valid regime, produces probabilistic forecasts of the expected normal behaviour, and compares the observed consumption with the forecast interval. Anomaly evidence is computed from the deviations beyond the forecast bounds, using a short temporal tolerance to absorb small timing shifts around sharp transitions. The daily decision separates pointwise diagnostics, daily severity, and context maintenance. Two severity measures are used: a soft reference severity, used to calibrate a robust threshold based on the interquartile range, and a decision severity, which aggregates only deviations that exceed the typical variability of the recent context. The protocol also includes in memory repair of previously anomalous points without overwriting raw observations, and a drift acceptance rule that updates the valid history boundary when persistent change becomes a new normal regime. The result is a reproducible and auditable methodology for daily anomaly detection in energy consumption, combining probabilistic forecasting, adaptive context selection, robust daily scoring, and explicit regime maintenance.