Intelligent online catastrophe assessment and preventive control via a stacked denoising autoencoder

Yangyang Liu,Mingyu Zhai,Jiahui Jin,Aibo Song,Jikeng Lin,Zhiang Wu,Yixin Zhao

DOI: https://doi.org/10.1016/j.neucom.2019.10.090

IF: 6

2020-01-01

Neurocomputing

Abstract:y In all areas of engineering, catastrophe assessment is an essential prerequisite for remedial action schemes. Modelers constantly push for more accurate models, and often meet goals by using increasingly complex, data mining-based blackbox models. However, system operators tend to favor interpretable models for after-the-fact preventive control (PC). While switching from blackbox to interpretable solutions, a tradeoffoccurs between accuracy and interpretability. To avoid this tradeoff, we develop an intelligent framework for online catastrophe assessment and PC via a blackbox stacked denoising autoencoder (SDAE) equipped with accuracy and the ability to derive a PC scheme. Specifically, we implement a transient stability cost-sensitive assessment (TSCA) and PC case in the context of a power grid. First, using only controllable variables, we build the TSCA model by adding a sigmoid unit on top of the SDAE. Considering power systems' conservatism, we explore a novel TSCA model's training criterion to determine the operation conditions' degrees of stability and divide them into three classes: stable, unstable, and boundary. Second, given an operation condition identified as unstable or boundary by TSCA model and its desired degree of stability, the PC model (the reverse of a TSCA model's mapping) consists of the top sigmoid's backward mapping and the stack of denoising decoders from trained SDAE. The former is formalized as an optimization problem to push back the desired degree of stability to a desired SDAE's highest-level abstraction. The latter decodes back the desired SDAE's highest-level abstraction to a desired operation condition (essentially, a PC scheme nearest to the controlled operation condition in the coordinates along the underlying causes that generate the observed data). This approach actually resembles operators' tendency to adjust and stabilize unstable conditions (in terms of underlying causes) with the fewest control actions. A simulation study on the IEEE New England 39-bus system shows that, as a blackbox technology, our framework not only provides superior online situational awareness, but also finds a viable PC scheme, thereby justifying its practicability in engineering. (C) 2019 Elsevier B.V. All rights reserved.