Power system analytics: anomaly detection, prediction, and mitigation
Loading...
Authors
Dwivedi, Anmol
Issue Date
2025-08
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Electrical engineering
Alternative Title
Abstract
Power transmission networks are critical infrastructures responsible for delivering electrical energyfrom generation sources to end-users. Traditionally, robust protection systems—including relay based
logic and automated safeguards—have successfully prevented many large-scale disruptions.
However, recent trends such as the extensive integration of renewable energy sources, growing
electricity demand, extreme weather events, aging infrastructure, and emerging cyber threats have
heightened the complexity and uncertainty faced by modern power grids. These factors significantly
elevate the risk of cascading failures and large-scale blackouts. Thus, the timely detection, accurate
prediction, and effective mitigation of such anomalies have become vital objectives for ensuring
power system reliability and resilience. To address these critical challenges, this dissertation
develops advanced, data-driven analytical methods rooted in statistical inference and graph machine
learning, tailored specifically for anomaly detection, localization, prediction, and mitigation. The first major focus of this dissertation addresses real-time detection and localization oftransmission line outages to improve grid reliability. Existing algorithms suffer from severe computational
bottlenecks, primarily due to simultaneously running multiple high-dimensional statistical
tests to detect and localize outages. This dissertation introduces a novel Graph-Guided Quickest
Change Detection (GG-QCD) framework that significantly reduces computational complexity
by decoupling outage detection from localization. Initially, a computationally lightweight one dimensional
spectral conformity metric tests the data’s conformity to the expected network structure
to rapidly detect an outage event. Upon detection, a localization stage, inspired by binary search
techniques, recursively partitions the network topology to efficiently pinpoint the affected transmission
line dramatically reducing the complexity from O(L) high-dimensional parallel tests to
O(log L) sequential tests, with each test also benefiting from significantly lower computational
complexity. Simulations on standard IEEE benchmarks demonstrate that the GG-QCD method
achieves considerable computational savings at the expense of only a modest increase in detection
delay, making it a practical and scalable solution for real-time power grid monitoring. The second contribution focuses on improving grid resilience by predicting cascading failures. Previous blackout analyses indicate that operators frequently fail to recognize incremental network changes, permitting hidden failures to accumulate until extensive disruptions occur. Existing predictive models generally struggle to capture concurrent spatio-temporal dependencies under dynamically evolving topologies, become computationally prohibitive at moderate network scales, and provide limited insight into causal relationships among network components. To addressthese issues, this dissertation introduces two complementary frameworks for cascading failure
prediction. The first framework formulates the prediction of risky fault chains as a partially
observableMarkov decision process (POMDP), approximately solved through a time-varying Graph
Recurrent Neural Network (GRNN) by employing a meta-reinforcement learning (RL) approach.
This model effectively captures the spatial and temporal dependencies inherent in power grids,
offering a scalable real-time predictive model. The second framework employs causal inference to
learn a directed latent graph that captures the cause-effect relationships among grid components,
providing a theoretical basis to quantify cascading interactions. Evaluations on IEEE benchmarks
demonstrate that both frameworks significantly enhance predictive accuracy and computational
efficiency by explicitly leveraging graph-structured representations and appropriately modeling
dependencies throughout the various cascading failure stages. The third and final contribution addresses cascading failure mitigation through sequentialremedial control actions. Large-scale failures often result from sustained congestion induced by
excessive load demands, gradually destabilizing the network. Traditional mitigation strategies
primarily emphasize bus-splitting techniques, often neglecting the potential of transmission line
disconnections and continuous generator control. To address this gap, this dissertation introduces a
novel physics-guided reinforcement learning (PG-RL) framework incorporating both discrete (line
reconnections/removals) and continuous (generator adjustments) actions within a unified hybrid
action space. By explicitly integrating sensitivity factors to guide RL exploration, the proposed
PG-RL framework substantially improves grid survival times compared to conventional black-box
RL methods. Empirical evaluations on the open-source Grid2Op platform illustrate that strategically
executed line removals can effectively delay or prevent cascading outages, highlighting previously
neglected remedial potentials. Additionally, the incorporation of continuous-valued generator
adjustments further enhances system resilience, offering a complementary alternative to traditional
bus-splitting approaches and paving the way for broader, more adaptive grid control strategies.
Description
August2025
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY