Akademik Veri Yönetim Sistemi
5th Mediterranean Conference on Pattern Recognition and Artificial Intelligence, MedPRAI 2021, Instanbul, Türkiye, 17 - 18 Aralık 2021, cilt.1543 CCIS, ss.315-328
© 2022, Springer Nature Switzerland AG.Smart grids find energy prices by comparing consumer demand with supply data. Since this is a time-sensitive process they need to predict smart grid stability dynamically. Power grid frequency rises in times of overproduction and decreases in times of underproduction. Using this feature of the grid, Decentral Smart Grid Control (DSGC) ties the grid frequency to energy price and gives us a mathematical model stability prediction. However, this solution comes with “fixed input” and “stability” issues. For solving this in a previous work we suggested using deep learning (DL) models. We compared multiple DL models, found one with 99.62% and showed DL models can give new insights to simulated grid stability prediction. However, since DL models are black boxes, the model lacked any information about why and how the system works. In general, this opaqueness of DL models stands in between them and wide-spread use in engineering. In this paper we used Layer-Wise Relevance Propagation (LRP) to find relevance scores of each input and to make our system human understandable. We show that the most important input in the DSGC system is reaction times of participants, followed by price elasticity coefficient and power consumption or generation have little to none effect on stability.