| Περίληψη: | The transition from obsolete Distribution Grids with centralized generation factories to a modern smart grid with an increasing Distributed Generation penetration has made DSOs desire a reliable grid protection system. Therefore, research of Distribution Grid surveillance algorithms has been motivated with the main goal of detecting faults in a smart grid, classifying their type and then pinpointing the fault’s location in the network for its immediate restoration. In this field Neural Networks are applied considerably because they dominate categorization problems and become more and more attractive due to their constantly decreasing decision-making times. Therefore, the purpose of the dissertation is to implement a complete application for Fault Location Identification and Classification (FLIC) of error affecting the healthy operation of a Low Voltage (LV) smart grid using Convolutional LSTMs for timeseries processing. Deploying a series of ANNs for Faulty Feeder and Branch Detection, Localization and Fault Type Classification and optimizing their hyperparameter using the Tree-structured Parzen Estimator(TPE) algorithm approach, accuracies of even 97% are reached. In addition, a simulated dataset composition algorithm is presented for LV grid fault timeseries measurements. This work’s target is the generalizability of the effort to all networks that meet certain specifications.
|
|---|
