Data-Driven Security Assessment of the Electric Power System

Abstract

The transition to a new low emission energy future results in a changing mix of generation and load types due to significant growth in renewable energy penetration and reduction in system inertia due to the exit of ageing fossil fuel power plants. This increases technical challenges for electrical grid planning and operation. This study introduces a new decomposition approach to account for the system security for short term planning using conventional machine learning tools. The immediate value of this work is that it provides extendable and computationally efficient guidelines for using supervised learning tools to assess first swing transient stability status. To provide an unbiased evaluation of the final model fit on the training dataset, the proposed approach was examined on a previously unseen test set. It distinguished stable and unstable cases in the test set accurately, with only 0.57% error, and showed a high precision in predicting the time of instability, with 6.8% error and mean absolute error as small as 0.0145.

Versatile and Robust Transient Stability Assessment via Instance Transfer Learning

“…The augmented dataset includes a same-distribution set of 10,000 samples and an auxiliary set of 5,000 samples, with the same proportion of classes in the two datasets. 2 To enable comparison with standard ML approaches, we also generate a set of labelled samples for training those data models. The training set includes 15,000 samples, where all loads in the system are randomly varied between 60% and 140% of their initial values.…”

“…In accordance with the conclusion of the two analyses, the loads at buses 4, 8, 15, 16, 20, 21, 23 and 24 are selected to produce the load scenarios of the auxiliary set. 2 The Sobol sampling approach is utilised throughout this study. eight varying loads (red arrows).…”

“…In these, data is collected from time-domain simulations, sampled from a fraction of the space of expected system states, and used off-line to train a data model. Such models, with promising results from both deep learning (DL) [2], [3] and conventional ML models [4], [5], can then be used to rapidly predict the stability of other system states. In addition, Transfer Learning, a collection of methods for improving the performance of data models on data with different distributions or feature spaces, has recently gained attention in power system contexts.…”