Load anomaly detection in power systems using modified CUSUM with future load profile nomination (FLPN) under hourly and sub-hourly resolution

Abstract

This paper investigates the use of Future Load Profile Nominations (FLPNs) as a forward-declared behavioral reference for load anomaly detection in power systems using a CUSUM-based framework. Although high-resolution monitoring improves observability, it also amplifies load variability, measurement noise, and non-stationary consumption behavior, particularly under sub-hourly resolution. Conventional anomaly detection approaches relying on historical baselines often struggle to distinguish genuine anomalies from legitimate demand fluctuations, leading to elevated false discovery rates. Rather than proposing a new detection algorithm, this study examines how reframing anomaly detection around intent-aware residuals improves detection stability across temporal resolutions. Three detection strategies are evaluated: traditional CUSUM with a forecasted baseline, traditional CUSUM with an FLPN reference, and a modified CUSUM with FLPN incorporating residual normalization and two-sided accumulation. Performance is assessed on the IEEE 14-bus system under progressively reduced spike-dip anomaly magnitudes using hourly and 5-minute sampling, with robustness further examined on the IEEE 57-bus system without parameter retuning. Results from 100 Monte Carlo simulations show that conventional CUSUM formulations degrade under temporal refinement, whereas the modified FLPN-assisted formulation maintains low false discovery rates and near-zero missed detections across the evaluated scenarios.