• DOI:10.1145/2821650.2821676
• Corpus ID: 2742485

Microgrid Losses: When the Whole Is Greater Than the Sum of Its Parts

@article{Buevich2016MicrogridLW,
  title={Microgrid Losses: When the Whole Is Greater Than the Sum of Its Parts},
  author={Maxim Buevich and Xiao Zhang and Oliver Shih and Dan Schnitzer and Tristan Escalada and Arthur Jacquiau-Chamski and Jon Thacker and Anthony G. Rowe},
  journal={2016 ACM/IEEE 7th International Conference on Cyber-Physical Systems (ICCPS)},
  year={2016},
  pages={1-10}
}

• Maxim Buevich, Xiao Zhang, +5 authors Anthony G. Rowe
• Published 4 November 2015
• Computer Science
• 2016 ACM/IEEE 7th International Conference on Cyber-Physical Systems (ICCPS)

Non-Technical Loss (NTL) represents a major challenge when providing reliable electrical service in developing countries, where it often accounts for 11-15% of total generation capacity [1]. NTL is caused by a variety of factors such as theft, unmetered homes, and inability to pay, which at volume can lead to system instability, grid failure, and major financial losses for providers. In this paper, we investigate error sources and techniques for separating NTL from total losses in microgrids. We adopt and compare two classes of approaches for detecting NTL: (1) model- driven and (2) data- driven. The model-driven class considers the primary sources of state uncertainty including line losses, meter consumption, meter calibration error, packet loss, and sample synchronization error. In the data-driven class, we use two approaches that learn grid state based on training data. The first approach uses a regression technique on an NTL-free period of grid operation to capture the relationship between state error and total consumption. The second approach uses an SVM trained on synthetic NTL data. Both classes of approaches can provide a confidence interval based on the amount of detected NTL. We experimentally evaluate and compare the approaches on wireless meter data collected from a 525-home microgrid deployed in Les Anglais, Haiti. We see that both are quite effective, but that the data-driven class is significantly easier to implement. In both cases, we are able to experimentally evaluate to what degree we can reliably separate NTL from total losses.