Machine Learning for Data-Driven Indicators Applied to Power Distribution System

Abstract

Indicators play an important role as they offer a quick overview of the system performance. However, obtain each indicator for each component of the power distribution system is cumbersome using classical approaches because the number of devices and data that must be studied is extensive. The main purpose of this work is to take advantage of machine learning algorithms to: (i) learn patterns from our data, and (ii) compute prediction-based indicators (we called data-driven indicators), that can be used to understand and improve distribution network performance. Our proposed methodology, used a long short-term memory (LSTM) auto-encoder architecture as a feature extractor in order to reduce the dimensionality, and then we used an LSTM forecaster network to make a daily prediction using smart-meters measurements. Finally, we employed the predicted values to compute the standardized indicators and ranked them based on the critical state. We carried out this analysis using real-world data collected at the State University of Campinas (UNICAMP). Our findings suggest that our proposed methodology can be suitable for power distribution networks where we faced with the problem of modeling unbalanced three-phase systems and with low X/R ratios.