EPE Association: EPE 2023 - DS3l: Application of Artificial Intelligence to Power Electronics and Drive Systems

EPE 2023 - DS3l: Application of Artificial Intelligence to Power Electronics and Drive Systems
You are here: EPE Documents > 01 - EPE & EPE ECCE Conference Proceedings > EPE 2023 ECCE Europe - Conference > EPE 2023 - Topic 10: Data Analysis, Artificial Intelligence and Communication > EPE 2023 - DS3l: Application of Artificial Intelligence to Power Electronics and Drive Systems

   [return to parent folder]

   Exploring the Effectiveness of Different State Spaces and Reward Functions in Reinforcement Learning-based Control of a DC/DC Buck Converter
By Parisa RANJBARAN, Shima SHAHNOOSHI, Javad EBRAHIMI, Alireza BAKHSHAI, Praveen JAIN [View]
[Download]

Abstract: The precise control of DC/DC buck converters iscritical for achieving stability, efficiency, andreliability of power systems. Reinforcementlearning (RL), a type of machine learning, hasshown promising results in solving controlproblems in various domains. The proposed studyaims to explore the potential of using RLtechniques, specifically the deep deterministicpolicy gradient (DDPG) algorithm, to controlDC/DC buck converters and to identify theoptimal combination of state space and rewardfunctions to achieve the best control performancefor DC/DC buck converters. This is done byanalyzing the performance of the control systemand comparing different sets of state space andreward functions. The findings of this study havethe potential to contribute to the development ofimproved control systems for power electronics.This research could ultimately lead to moreefficient, stable, and reliable power systems.

   Immune Neuro-Fuzzy Network Based System for Collision Free Motion Control of Unmanned Electrical Vehicles
By Anna BEINAROVICA, Mikhail GOROBETZ, Leonids RIBICKIS [View]
[Download]

Abstract: This paper presents current results of the research, aimed at the developing of motion control algorithms and systems for autonomous electric vehicles , and discusses one of the goals of the research - collision probability minimization and traffic flow optimization. The task of the current research is to develop a computer model and a laboratory prototype, which would analyze the input data, calculate and minimize the collision probability by changing motion parameters and optimize traffic flow. Computer simulations prove the workability and advantages of the developed model.

   LSTM Data-Driven Model of Multi-scene Virtual Synchronous Generator
By Jiangbin TIAN, Jinbin ZHAO, Guohui ZENG, Xiangchen ZHU, Zhenhua ZHANG, Yuzong WANG [View]
[Download]

Abstract: This paper proposes a data-driven modeling approach using LSTM to accurately describe the dynamic characteristics of Virtual Synchronous Generator (VSG) systems. VSG technology allows grid-connected inverters to resemble synchronous generators externally. However, the commonly used small-signal model faces challenges in capturing the complexity of VSG systems, particularly in complex scenarios. The proposed LSTM-based data-driven model considers the impact of irrational factors, providing accurate and stable VSG system modeling. Experimental results demonstrate the superiority of the LSTM neural network-based data-driven VSG model over the small-signal model and typical data-driven models in terms of accuracy and stability.

   Reinforcement Learning-based Control of a Buck Converter: A Comparative Study of DQN and DDPG Algorithms
By Shima SHAHNOOSHI, Parisa RANJBARAN, Javad EBRAHIMI, Alireza BAKHSHAI, Praveen JAIN [View]
[Download]

Abstract: Reinforcement learning (RL) has emerged as apromising approach for controlling powerelectronics systems due to its ability to controlnonlinear systems without accurate models, andhandle unknown changes. This paper presents acomparative study of two widely used deep RLalgorithms, Deep Deterministic Policy Gradient(DDPG) and Deep Q-Network (DQN), forcontrolling the output voltage of DC/DC buckconverters. DDPG and DQN are tailored tohandle continuous and discrete environments,respectively. The performance of thesealgorithms is evaluated through simulations withsteady-state error, overshoot, settling time, andtraining time serving as evaluation metrics. Thesimulation results demonstrate that both DDPGand DQN show excellent performance incontrolling power converters, with DDPGgenerally performing better than DQN in systemswith unpredictable changes.

EPE 2023 - DS3l: Application of Artificial Intelligence to Power Electronics and Drive Systems
You are here: EPE Documents > 01 - EPE & EPE ECCE Conference Proceedings > EPE 2023 ECCE Europe - Conference > EPE 2023 - Topic 10: Data Analysis, Artificial Intelligence and Communication > EPE 2023 - DS3l: Application of Artificial Intelligence to Power Electronics and Drive Systems
	[return to parent folder]

	Exploring the Effectiveness of Different State Spaces and Reward Functions in Reinforcement Learning-based Control of a DC/DC Buck Converter By Parisa RANJBARAN, Shima SHAHNOOSHI, Javad EBRAHIMI, Alireza BAKHSHAI, Praveen JAIN	[View] [Download]
Abstract: The precise control of DC/DC buck converters iscritical for achieving stability, efficiency, andreliability of power systems. Reinforcementlearning (RL), a type of machine learning, hasshown promising results in solving controlproblems in various domains. The proposed studyaims to explore the potential of using RLtechniques, specifically the deep deterministicpolicy gradient (DDPG) algorithm, to controlDC/DC buck converters and to identify theoptimal combination of state space and rewardfunctions to achieve the best control performancefor DC/DC buck converters. This is done byanalyzing the performance of the control systemand comparing different sets of state space andreward functions. The findings of this study havethe potential to contribute to the development ofimproved control systems for power electronics.This research could ultimately lead to moreefficient, stable, and reliable power systems.

	Immune Neuro-Fuzzy Network Based System for Collision Free Motion Control of Unmanned Electrical Vehicles By Anna BEINAROVICA, Mikhail GOROBETZ, Leonids RIBICKIS	[View] [Download]
Abstract: This paper presents current results of the research, aimed at the developing of motion control algorithms and systems for autonomous electric vehicles , and discusses one of the goals of the research - collision probability minimization and traffic flow optimization. The task of the current research is to develop a computer model and a laboratory prototype, which would analyze the input data, calculate and minimize the collision probability by changing motion parameters and optimize traffic flow. Computer simulations prove the workability and advantages of the developed model.

	LSTM Data-Driven Model of Multi-scene Virtual Synchronous Generator By Jiangbin TIAN, Jinbin ZHAO, Guohui ZENG, Xiangchen ZHU, Zhenhua ZHANG, Yuzong WANG	[View] [Download]
Abstract: This paper proposes a data-driven modeling approach using LSTM to accurately describe the dynamic characteristics of Virtual Synchronous Generator (VSG) systems. VSG technology allows grid-connected inverters to resemble synchronous generators externally. However, the commonly used small-signal model faces challenges in capturing the complexity of VSG systems, particularly in complex scenarios. The proposed LSTM-based data-driven model considers the impact of irrational factors, providing accurate and stable VSG system modeling. Experimental results demonstrate the superiority of the LSTM neural network-based data-driven VSG model over the small-signal model and typical data-driven models in terms of accuracy and stability.

	Reinforcement Learning-based Control of a Buck Converter: A Comparative Study of DQN and DDPG Algorithms By Shima SHAHNOOSHI, Parisa RANJBARAN, Javad EBRAHIMI, Alireza BAKHSHAI, Praveen JAIN	[View] [Download]
Abstract: Reinforcement learning (RL) has emerged as apromising approach for controlling powerelectronics systems due to its ability to controlnonlinear systems without accurate models, andhandle unknown changes. This paper presents acomparative study of two widely used deep RLalgorithms, Deep Deterministic Policy Gradient(DDPG) and Deep Q-Network (DQN), forcontrolling the output voltage of DC/DC buckconverters. DDPG and DQN are tailored tohandle continuous and discrete environments,respectively. The performance of thesealgorithms is evaluated through simulations withsteady-state error, overshoot, settling time, andtraining time serving as evaluation metrics. Thesimulation results demonstrate that both DDPGand DQN show excellent performance incontrolling power converters, with DDPGgenerally performing better than DQN in systemswith unpredictable changes.