Expert System for Real-Time Anomaly Identification and Learning Power System Disturbances

Abstract

The urgent need to decrease global carbon emissions to meet the Paris Climate Agreement calls for sustainable methods to manage growing electrical demands. This energy transition requires the decommissioning of large fuel-based power plants and simultaneously replacing them with power-electronics interfaced intermittent renewable energy sources and loads. These developments pose high stress on our ageing grid infrastructure, leading to an increased level of unanticipated electrical disturbances that, if left unchecked, might lead to total grid collapse. The thesis presents an expert system that fortifies the ever-evolving grid with advanced event identification and learning architecture engineered to protect against contemporary and evolving grid disturbances. The chosen design perspectives are intended to assure trust in academic algorithms and bridge the expanding gap between the academia-industry. In this context, the dissertation has three main contributions namely:
♦ A real-time PMU-based distribution state estimation.
♦ A near-real-time dynamic incremental learning-based event classifier.
♦ An adaptive human-in-the-loop event identification methodology.

First, to conduct extensive simulations on variety of model-driven and data-driven algorithms, a close to real-life simulation environment needs to be set-up. Using RTDS a cyber-physical replica of a 50 kV ring network operated by Stedin B.V. in the Zeeland area of the Netherlands is developed. Further, the grid is upgraded in 3 operational stages to meet steady-state, quasi-steady-state and dynamic-state conditions. This forms the benchmark grid for all further studies. Subsequently, as a first step towards real-time grid situational awareness, state-of-the-art EKF- and UKF-based state estimation algorithms are developed, tested and validated to achieve complete grid observability in terms of determined node voltage phasors for the grid. With enough confidence in terms of SE accuracy and computational efficiency in the steady-state, the PMU-based state estimator increases complexity by QSS operation and finally, by adopting an anomaly detection, discrimination, and identification module, the PMU-based state estimator is enhanced to co-simulate within the fast refresh rates of PMUs under a fully dynamic grid with abrupt SLC and multiple bad-data events.

Second, with PMU-detectable events addressed, events with complex temporal signatures are systematically identified using data-driven models. Recommendations are developed for a forecast-based event detection model and subsequent real-time data pre-processing, which collect disturbance signatures. A multivariate 1D CNN classification model is designed to identify event types using disturbance signatures in real time. In the first stage, simulations are performed for events which are known and previously trained by the model. In the next stage, the DIL strategy is used to adapt the data-driven model for unforeseen and statistically drifted event types. The classification accuracy, memory consumption, and computational efficiency are used as performance metrics to validate in near-real-time conditions.

Third, in order for data-driven models to meet industrial expectations, an AdInFier expert system is developed, which primarily adds a validation stage to verify the classification results using an unsupervised learning approach. A Soft-DTW technique is used for event representatives that will be compared with incoming disturbance signatures to provide a similarity score. The classifier-validator duo provides a two-stage approach for event identification so that control actions can be actuated in real time in high-stakes environments of control centres. Subsequently, we inculcate a human-in-the-loop approach within an AI environment to deal with complex, contradictory situations where the grid collected data is not mature enough for models to decide on the event type. This step is mainly to add domain expert knowledge in the solutions of over-deterministic data-driven models.

The main purpose of this dissertation is to get a step closer to real-life implementation of state-of-the-art model-driven algorithms and ensure trust in the new cutting-edge data-driven domains with the ultimate goal of meeting industrial requirements. As future recommendations, we propose further enhancements to the AdInFier expert system in terms of control actions and solution fulfilment capabilities, so that we can safely manoeuvre in today's fast-paced technological landscape.