IRUS Total

Active network management for improved flexibility in distribution networks

File Description SizeFormat 
PerezOlvera-J-2021-PhD-Thesis.pdfThesis18.86 MBAdobe PDFView/Open
Title: Active network management for improved flexibility in distribution networks
Authors: Perez Olvera, Julio
Item Type: Thesis or dissertation
Abstract: Decarbonisation efforts have prompted the rapid uptake of low-carbon technologies (LCTs) in distribution networks. Their power use and delivery are highly variable and can cause various voltage-regulation and thermal-overload problems, thus making it difficult to maintain a secure operation of these networks. The problem is compounded by the lack of instrumentation in low-voltage networks. Two specific problems are identified: 1) the need for greater flexibility in power routing in distribution networks and 2) the need for optimising controllers operating with partial visibility of the network state and with practical computational efficiency. Power flow flexibility is provided by AC/DC power electronic converters and DC links, as part of an Active Network Management (ANM) scheme. An Optimal Power Flow (OPF) algorithm is modified to include DC links and serves as a benchmark for dispatching the DC power flows but is deemed too computationally intensive for real-time control. Alternatives to the OPF, based on regression and reinforcement learning, were examined using casestudy networks. Regression mapping of the OPF output is proposed to move the computational effort offline and to address the partial observability of the network. However, it was found to have poor accuracy and a limited ability to generalise. For better generalisation and accuracy, a model-free reinforcement learning method based on deep Q networks was proposed. The performance was on a par with the OPF while being able to generalise to different loading conditions, however the training required several hours. A model-based algorithm using an explicit model of the network was proposed to further improve the accuracy and reduce the computational burden. The training time was in the range of minutes with good accuracy compared to the OPF. ANM using power converters together with controllers based on reinforcement learning represent a compelling alternative to further integrate LCTs in distribution networks.
Content Version: Open Access
Issue Date: Sep-2020
Date Awarded: Apr-2021
URI: http://hdl.handle.net/10044/1/94477
DOI: https://doi.org/10.25560/94477
Copyright Statement: Creative Commons Attribution NonCommercial NoDerivatives Licence
Supervisor: Green, Timothy
Sponsor/Funder: Mexican Energy Ministry (SENER)
Mexican National Council for Science and Technology
Funder's Grant Number: PhD Scholarship number: 441647
Department: Electrical and Electronic Engineering
Publisher: Imperial College London
Qualification Level: Doctoral
Qualification Name: Doctor of Philosophy (PhD)
Appears in Collections:Electrical and Electronic Engineering PhD theses

This item is licensed under a Creative Commons License Creative Commons