Evolutionary Optimization of a One-Class Classification System for Faults Recognition in Smart Grids

摘要

The Computational Intelligence paradigm has proven to be a useful approach when facing problems related to Smart Grids (SG). The modern SG systems are equipped with Smart Sensors scattered in the real-world power distribution lines that are able to take a fine-grained picture of the actual power grid state gathering a huge amount of heterogeneous data. Modeling and predicting general faults instances by means of processing structured patterns of faults data coming from Smart Sensors is a very challenging task. This paper deals with the problem of faults modeling and recognition on MV feeders in the real-world Smart Grid system that feeds the city of Rome, Italy. The faults recognition problem is faced by means of a One-Class classifier based on a modified k-means algorithm trained through an evolutive approach. Due to the nature of the specific data-driven problem at hand, a custom weighted dissimilarity measure designed to cope with mixed data type like numerical data, Time Series and categorical data is adopted. For the latter a Semantic Distance (SD) is proposed, capable to grasp semantical information from clustered data. A genetic algorithm is in charge to optimize system's performance. Tests were performed on data gathered over three years by ACEA Distribuzione S.p.A., the company that manages the power grid of Rome.