A Machine-Learning Pipeline for Large-Scale Power-Quality Forecasting in the Mexican Distribution Grid

Abstract

Electric power distribution networks face increasing factors for power-quality (PQ) deterioration, such as distributed, renewable-energy generation units and countless high-end electronic devices loaded as controllers or in standalone mode. Consequently, government regulations are issued worldwide to set up strict PQ distribution standards; the distribution grids must comply with those regulations. This situation drives research towards PQ forecasting as a crucial part of early-warning systems. However, most of the approaches in the literature disregard the big-data nature of the problem by working on small datasets. These datasets come from short-scale off-grid configurations or selected portions of a larger power grid. This article addresses a study case from a region-sized state-owned Mexican distribution grid, where the company must preserve essential PQ standards in approximately 700 distribution circuits and 150 quality-control nodes. We implemented a machine-learning pipeline with nearly 4000 univariate forecasting models to address this challenge. The system executes a weekly forecasting pipeline and daily data ingestion and preprocessing pipeline, processing massive amounts of data ingested. The implemented system, MIRD (an acronym for Monitoreo Inteligente de Redes de Distribution—Intelligent Monitoring of Distribution Networks), is an unprecedented effort in the production, deployment, and continuous use of forecasting models for PQ indices monitoring. To the extent of the authors’ best knowledge, there is no similar work of this type in any other Latin-American distribution grid.

Keywords

• steady-state power quality indices
• time-series forecasting
• machine learning
• artificial neural networks