Application of spatio-temporal data-driven and machine learning algorithms for security assessment

This document reports on the recent advancements on spatio-temporal data-driven and machine learning methods for static and dynamic security assessment, and their particular use cases.

Various data-driven methods can analyse the data collected within Wide-Area Monitoring Systems (WAMS) from phasor measurement units (PMUs). Data-driven methods are particularly promising to identify the dynamics of the current system from these measurements and to predict dynamic stability as dynamic oscillations can be dangerous for stability. Linear signal processing techniques can identify modal information and the current system behaviour via linear ringdown analysis. Beyond linear system analysis, the Koopman mode decomposition analyses nonlinear modes and dictionary analysis considers possible dynamic basis functions. Unsupervised learning can identify from the observed data the coherency across generators to identify the generators that respond dynamically similarly. To assess the stability in an early warning and early prevention system can analyse the various stability phenomena in parallel to save computational time which is key.

Machine learning (ML) methods are promising for applications around near-real-time dynamic security assessment (DSA) and security assessment (SA) as they produce instantaneously the security label, often based on convolutional neural networks (CNNs). Such ML-based SA workflows are promising for low-inertia systems as the timeframes of dynamic events are shorter and the security boundaries become easier separable with ML. To reduce the computational times of ML workflows further, feature extraction and dimensional reductions have been explored. Beyond security assessment, ML methods can also be used for preventive control and event detection. Using ML for preventive control, the total transfer capability can be predicted for preventive operational planning, and deep reinforcement learning can sequentially decide under emergencies. Additionally, analysing the spectrum from synchrophasor data, then using the dynamic wavelet transform can be combined with CNN-based classification to detect (and classify) an event.

Several applications of data-driven approaches were developed ranging from Python toolboxes and web-based applications for ring down analysis, coherency identification and the identification of dynamical parameters over the analysis on real-power systems such as on the Chilean, Ecuadorian, Japanese’s and Swedish systems. The Python toolbox for ringdown oscillations analysis single signals or a set of registered signals from events measured in electrical power systems, which can be synchrophasor data, containing the same time stamp. The Python toolbox for coherency identification uses a model-view controller and Django Framework in the cloud, and the web-based application for identification of the dynamic parameters for reduced orders with a contingency analysis demonstrated on Chilean power system. The Functional Basis Analysis (FBA) techniques was applied to both static and dynamic phasor-based methods using synthetic and real waveform, showing benefits for various faults. The Koopman decomposition were applied to Japan’s power system, there based on measured frequency, the dominant Koopman eigenvalues and modes are extracted. Finally, an early warning and prevention method was developed for preventing blackouts in Swedish system, in response to a blackout from 2003 triggered by aperiodic small-signal rotor angle stability (ASSRA).

ML approaches are applied to various applications and tested on several systems. The applications range from static security assessment with stratified cross-validations and estimating the global operating system state over predicting the time-domain trajectories and assessing the transients stability in real-time, then detecting events, and evaluating real-time synchrophasor data, subsequently, the preventive and secure control with hybrid deep learning. Within these applications, the use of ML is demonstrated on systems ranging from IEEE 39, 68, and 140 bus systems, or real power systems such as the Brazilian, and interestingly the case for using ML gets stronger in low-inertia power systems than in high-inertia systems.