Data-Driven Voltage Control of DERs Integrated Distribution Grids Using Deep Reinforcement Learning
Loading...
Authors
Hossain, Rakib
Issue Date
2024
Type
Dissertation
Language
en_US
Keywords
Alternative Title
Abstract
The proliferation of Distributed Energy Resources (DERs) in modern power distribution grids presents a significant challenge for voltage control and power loss due to their intermittent nature and diverse operational characteristics. Although different control strategies have been developed to improve the voltage profile of the distribution networks, Volt-VAR control (VVC) is one of the efficient tools that can be used to control the reactive power set-points of legacy reactive power resources or newly added inverters to regulate the voltage of distribution networks. VVC methods regulate the voltage by exploiting the reactive power absorption/injection capabilities of capacitor banks, and smart inverters. Although existing classical VVC control/ optimization based approaches can control the set-point of the slow responding legacy devices and fast responding smart inverters to control the the voltage of the distribution network, machine Learning (ML) based VVC/VVO approaches especially deep reinforcement learning (DRL) based VVC approaches has become popular for real time control of the voltage. DRL can optimally control the reactive power set-point of the inverter to minimize the voltage fluctuation and avoid the voltage violation. This dissertation investigates the application of Deep Reinforcement Learning (DRL) techniques for voltage control in distribution grids with high penetration of DERs. Leveraging the capabilities of artificial intelligence, specifically deep neural networks and reinforcement learning algorithms, our approach addresses the complexity of managing voltage levels in the presence of decentralized energy sources. It presents a comprehensive analysis of the challenges posed by high DER penetration and demonstrate the effectiveness of DRL-based control strategies in optimizing grid voltage profiles. By training deep reinforcement learning agents on real-time data from distribution networks with a diverse array of DERs, we achieve precise and adaptive voltage control. The proposed methodology not only ensures the stability and reliability of the distribution grid under varying operating conditions but also maximizes the utilization of renewable energy sources. Existing DRL based approaches do not consider the topology information of the network and they don't consider the consequence of missing/unobservable data. Therefore, incorporating graph convolution layers and generative adversarial network with DRL algorithm, this research address these problems. The implementation of the existing DRL based approaches in real power grids has been impeded by the absence of clear assurances regarding stability and safety. Therefore, this research utilizes a monotonically decreased neural network that can guarantee the stability of the policy during training and execution processes. The capability of the DRL based approaches in real-time implementation in tested using real-time digital simulator (RTDS). Through extensive simulations and comparative analyses, we showcase the superior performance of our DRL-based voltage control approach in comparison to traditional methods. The results underscore the potential of advanced machine learning techniques in enhancing the reliability and efficiency of distribution grids amidst the rapid integration of DERs. This research contributes valuable insights and practical solutions for utilities and researchers aiming to optimize voltage control strategies in distribution systems with a high penetration of renewable energy resources.