نوع مقاله : پژوهشی
تازه های تحقیق
عنوان مقاله English
نویسندگان English
Unplanned power outages disrupt grid stability and increase operational costs, posing a major threat to power system efficiency. In this paper, we propose a robust outage prediction model that combines an Autoencoder-based feature extractor with a residual multi-layer perceptron (MLP) classifier. The novelty of our approach lies in its ability to maintain high predictive performance while eliminating reliance on geographic features such as latitude and longitude—commonly required by traditional models. We first train the Autoencoder on a rich, unlabeled dataset of weather and energy demand data collected over two decades (2000–2024) across Maryland, USA. The learned latent representations are then used to augment a supervised classification model trained on labeled outage data. Our final model achieves an F1-score of 81% even without location-based features, compared to 90% when using all features. This generalizability enables the deployment of predictive tools in previously unseen regions, directly enhancing grid flexibility, reliability, and system efficiency.
کلیدواژهها English
Copyright © Reihane Montazeri Najafabadi, Mohammadreza Shams
License
This article is released under the Creative Commons Attribution (CC BY 4.0) license. Anyone is free to copy, share, translate, and adapt this article for any purpose, whether commercial or non-commercial, as long as proper citation is given to the authors and original publication.
