Binary Logistic Regression and Cross-Entropy Techniques Applied to Composite Reliability Evaluation of Power Systems

The reliability evaluation of composite generation and transmission systems is a problem with high computational effort. The optimization of algorithms that carry out this assessment has historically been deeply studied, with emphasis on two approaches: (i) reducing the sample space necessary to represent the event of interest via variance reduction techniques (VRTs); and (ii) minimizing the computational effort of state adequacy analyses via machine learning classifiers. Although they have already been used together, no work has proposed using the knowledge produced by VRTs to improve the classifier performance. In this paper, a quasi-sequential Monte Carlo simulation is proposed, optimized by cross-entropy and binary logistic regression techniques, in which the VRT output information is passed on to the classifier to improve the performance of the explanatory variables related to system failures. To validate the proposed method, several evaluations are carried out with a modified version of the IEEE RTS-1996, with characteristics similar to those of modern interconnected systems.