Collective Continuity Index prediction through time with deep learning and auto-regressive models

This project was developed in 2018 by Lucas Thimoteo and Melissa Costa as part of the final credits to receive their Bachelor Degree in Electrical Engineering by Rio de Janeiro State University

The main goal was to develop a deep learning model to make a multistep time series prediction of the collective continuity index of an electrical distribution company.

Abstract:

Throughout 20th century, most of activities in the industry, commercial and residential sectors became more dependent on electricity. To ensure distribution company’s quality of service to consumers, the Brazilian National Agency of Electrical Energy stipulates annual limits for collective continuity indicators on every concessionaire tariff revision. On the other hand, concessionaires seek to offer the best service at the least cost. Therefore, it is in the interest of these companies to understand the variables that affect quality of service, as well as monitor collective continuity indicators levels and also execute optimized improvements in the system, to mitigate revenue losses due to electricity supply interruption. This work aims to study the historical behavior of collective continuity indicators and propose models to predict their future values, within an one year horizon, so that this prediction can ancilliate the decision making process when planning the distribution system. Utilized models consisted on Artificial Neural Networks and also ARIMA statistical models. Input data consisted on monthly values of Light SESA collective indicators. This company is responsible for the concession of electricity distribution on most of the metropolitan region of Rio de Janeiro state. Predictions were compared to historical data through the root mean squared error and the median absolute percentage error. Both models obtained approximately 40% of median absolute percentage error and 0,3 of root mean squared error. However, the ARIMA model shows slightly superior results in some occasions.

Keywords: Interruption Duration Equivalent Index. Interruption Frequency Equivalent Index. Artificial Neural Networks. Time Series