State Estimation for Unobservable Distribution Systems via Deep Neural Networks
Collaborators: Kursat Rasim Mestav
Advisor: Prof. Lang Tong
Cornell University
Sponsor: Iberdrola Foundation
Award: IEEE PESGM Best Conference Paper
Abstract: The present distribution systems are not well metered and in general unobservable. To unlock the full potential of Distributed Energy Resources (DERs), a modernization of the distribution system is necessary to provide tighter control of power flow in real-time operations, which requires effective state estimation. An essential barrier to state estimation for real-time control is unobservability. Although smart meters at the edge of the network have been deployed progressively, these type of measurements are typically at a much slower time scale incompatible with the more rapid changes of DER such as solar generation. Realizing state estimation for real-time operation in distribution systems, therefore, requires a fundamentally different approach from that used in transmission systems one that overcomes the difficulty of lack of measurements.
A Bayesian approach is proposed that combines Bayesian inference with a deep neural network to achieve the minimum mean squared error estimation of network states for real-time applications. The proposed technique learns probability distributions of net injection from smart meter data and generate samples for training a deep neural network. Results show that the proposed technique offers significant improvement in estimation accuracy and computation cost over weighted least squares methods with pseudo-measurements. Simulations are also used to evaluate robustness of the proposed Bayesian method against estimation errors in distribution learning and bad data.