Data analysis and management for optimal application of an advanced ML-based fault location algorithm for low voltage grids

Abstract

As the need for automatization of the electricity grid’s fault diagnosis schemes is rising, the application of technologies such as the artificial intelligence (AI) can provide practical solutions to the problem. AI can overcome the challenges that complex topologies like those of the low voltage (LV) smart grids pose and prove to be a powerful tool in the development of advanced fault diagnosis methods. An important parameter for the success of any AI-based method is the quality of data. Therefore, in this paper a data analysis is performed in order to evaluate the type of data produced by a small LV grid and an representative AI algorithm’s response to those. In the context of this analysis, the most important features and meters were identified. Furthermore, as a response to the large volume of available data, a data management strategy is proposed. The strategy combines original and reshaped features. For this purpose, five dimensionality reduction methods are tested and compared. Truncated-SVD is deemed the most appropriate and is subsequently utilized for the reshaping of the dataset that is introduced to the XGBoost fault location model. The integration of the dimensionality reduction technique in the algorithm results in the decrease of the computational time and the dataset’s size and in a higher generalizability of the algorithm. Thus, the application of the proposed method is not limited by the grid’s topology. The method’s robustness was verified against various influencing parameters such as the fault resistance, the size of the dataset, the loss of data and the photovoltaics’ penetration level. The overall algorithm achieved a mean squared error of 13.26 and a training and test accuracy of more than 99% when tested on the CIGRE LV benchmark grid.