Training autoencoders for state estimation in smart grids

• Main Author: Rui Miguel Machado Oliveira
• Publication Date: 2017
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: https://hdl.handle.net/10216/106291
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Since its massive deploy in power systems' control centres, the state estimator is regarded as a tool of critical importance due to its capability of determining a system operating point with remarkable precision. Indeed, one of the essential stages to obtain that operating point is identifying the system topology, i.e., the open or closed status of the grid's switches and breakers. As a result, a considerable amount of methodologies was proposed with the purpose of detecting the correct topology. First efforts related to this subject date back to the 80's and preconized the use of strict mathematical equations and, subsequently, the analysis of their results to evaluate possible topology errors. However, other techniques showed the advantages of the application of artificial neural networks to this particular problem; among them, figure the fast and reliable results and immunity to convergence problems of mathematical, analytic methods. Hence, this thesis presents a novel perspective of the application of neural networks, deep learning and Information Theoretic Learning (ITL) to the problem of topology determination and, ultimately, to state estimation. Blending these concepts, the main objective is to predict the topology of a given power system based on pattern recognition of analog measurements, pointing, at the same time, alternative approaches on neural networks' training and architecture as well as computational tools capable of introducing enhanced efficiency to that process. With more detail, two different applications with elements of supervised and unsupervised training as well as with autoencoder architectures and typical feedforward neural networks will be explored to discover the topology of an IEEE RTS 24 test case. Apart from comparisons between their efficacy, other major point will be the contrasts between running times achieved by CPU and GPU computation, showing that the large spectrum of application of this last technique comprises also power systems.