Time Series-Analysis Based Engineering of High-Dimensional Wide-Area Stability Indices for Machine Learning

Raoult Teukam Dabou,C Y Chung,Innocent Kamwa,Chuma Francis Mugombozi

doi:10.1109/access.2021.3099459

Raoult Teukam Dabou, C Y Chung + Show 2 more

Open Access

https://doi.org/10.1109/access.2021.3099459

Copy DOI

Abstract

Information representative of actual power system dynamics is usually buried in masses of phasor measurement unit (PMU) data. To take full advantage of these data in early anticipation of stability loss, we propose to implement the high dimensional stability index (HDSI). This method allows the extraction of more than 500-labeled attributes describing generator response signals, such as speed and rate of change of transient energy function (RoCoTE). A combined 31 functions are computed from spectrum analysis based on the Periodogram and Welch methods, Lyapunov exponents, and wavelet transform approaches. The test databases are built by simulating faults on each line in the IEEE 39- and 68-bus networks. Applying comparative time-series analysis to such signal responses to disturbances then highlights the texture matrix of the stability attributes. A 10-fold support vector machine (SVM) is used to implement a HDSI-based stability prediction model, with its performance then compared to the artificial neural network (ANN), decision trees (DT), random forest (RF), and adaptive boosting (AdaBoost) models available in the statistical package R. While most methods performed similarly, with ~100% accuracy on test cases using the same set of HDSI-based attributes, the RF classifier with its associated Gini feature importance allows for explicit feature ranking and interpretation, which results in prioritization of frequency-domain over time-domain features.

Highlights

Time-series of post-disturbance response signals contain dormant and non-transparent information that is useful for understanding the dynamics of power systems in real time [1]
In light of the above literature, the present paper focuses on the combined use of Lyapunov exponents, wavelet and Fourier transforms, and spectral energy using periodograms and Welch methods to extract stability precursor attributes from time-series of generators rotor speeds and rate of change of transient energy function (RoCoTE) derived from Dynamic State Estimation (DSE) utilizing phasor measurement unit (PMU) data [1]
By combining the two datasets, with the first dominated by transient stability phenomena while the second is dominated by small-signal stability type of events [34], it is possible to derive a single machine-learning based stability predictor, able to generalize the stability features to all types of network dynamics; this leads to a single model that can work potentially for never seen networks or novel operating conditions on previously seen networks

Summary

Introduction

Time-series of post-disturbance response signals contain dormant and non-transparent information that is useful for understanding the dynamics of power systems in real time [1]. As pointed out in [6], this opacity is an important limitation to the maintenance of such models and their adoption by the industry In this context, the characterization of temporal responses prior to machine learning by physically interpretable attributes, via notions such as the maximum of spectral energy density or maximum of Lyapunov exponents, becomes an interesting alternative. The central tenet of the proposed approach is that by increasing drastically the type and number of physically interpretable ‘‘PMU data based catastrophic indicators’’ [17] involved in the machine learning (i.e. moving from a low dimensional to high-dimensional feature space), we should be able to improve performance and transparency of the power system stability prediction model. HDSI will simultaneously enhance classification performance of existing machine learning and increase the transparency and readability of the underlying algorithms decision rules

Objectives

Results

Discussion

Conclusion