Convolutional neural nets with hyperparameter optimization and feature importance for power system static security assessment

Abstract

Static security assessment (SSA) is fundamental in electrical network analysis. However, the growing complexity and variability of grid's operating conditions can make it tedious, slow, computationally intensive, and limited or impractical for on-line applications when traditional approaches are considered. Since this may hinder the emerging analytical duties of system operators, data-driven alternatives are required for faster and sophisticated decision-making. Although different machine learning algorithms (MLAs) could be applied, Convolutional Neural Networks (CNNs) are one of the most powerful models used in many advanced technological developments due to their remarkable capability to identify meaningful patterns in challenging and complex data sets. According to this, a CNN based approach for fast SSA of power systems with N-1 contingency is presented in this paper. To contribute to the automation of model building and tuning, a settings-free strategy to optimize a set of hyperparameters is adopted. Besides, permutation feature importance is considered to identify only a subset of key features and reduce the initial input space. To illustrate the application of the proposed approach, the simulation model of a practical grid in Mexico is used. The superior performance of the CNN alternative is demonstrated by comparing it with two popular MLAs.