Desarrollo de modelos de pronósticos mediante el uso de aprendizaje automático para la caracterización de desconexiones eléctricas que ocurren en la ciudad de estudio

Abstract

This study presents a machine-learning approach to characterize the origin of electrical outages in an urban area on Ecuador’s coastal region. The objective is to classify events into internally scheduled, internally unscheduled, externally scheduled, and externally unscheduled categories and to identify spatial patterns and associated factors; the hypothesis is that combining operational, network, temporal, climatic, and socio-economic variables enables accurate discrimination, justified by its usefulness for distribution-system management. Historical outage records from a power-distribution utility for 2023 and 2024 were compiled, and twelve predictive variables were integrated. Supervised classifiers, Random Forest and a multilayer perceptron, were trained with stratified validation and precision, recall, and F1 metrics; georeferencing was performed for the study city. Random Forest achieved an overall accuracy close to 99 %, and the MLP around 99.7 %, with higher performance for classes with greater support. The results confirm that the taxonomy of outage origin is separable using the twelve variables considered, and that geospatial visualization and prioritization by substation and sector provide useful inputs for planning and mitigation. Keywords: Machine Learning; Distribution System; Electrical Outages; Georeferencing; Coastal Ecuador.