To reduce fossil fuel consumption, which causes carbon dioxide emissions and global warming, renewable energy is gaining popularity. Among various renewable energy sources, wind energy is one of the most cost-effective ways to generate electricity. Numerous studies have been conducted to improve the performance of wind energy conversion systems (WECS) in various aspects. However, traditional control strategies employed in WECS often lead to lower efficiency, complicated implementation, complex system modeling, sophisticated drive circuit design, and suboptimal responses. This PhD thesis presents a comprehensive exploration of cutting-edge techniques for optimizing wind energy conversion systems, unified by the application of a proposed multi-agent reinforcement learning (MARL) method. The research is structured around three primary objectives, each contributing to the advancement of renewable energy technologies through the innovative use of MARL. Firstly, the thesis delves into the control of a neutral point clamped (NPC) power converter employed in a direct-drive permanent magnet synchronous generator (PMSG)-based WECS. The focus is on enhancing power quality and meeting grid code requirements for total harmonic distortion (THD). Traditional controllers like PI often struggle with parameter tuning and adaptability to varying operating conditions, resulting in suboptimal performance under dynamic and unbalanced scenarios. AI-based approaches, while more adaptive, typically require extensive offline training and detailed system modeling, making them less practical for real-time applications. The proposed approach eliminates the need for offline training and extensive system modeling, distinguishing itself from traditional machine learning (ML), neural network-based techniques, and PI-based methods. Through simulations and comparative analysis, the effectiveness of the MARL strategy is validated, particularly in handling unbalanced voltage sag scenarios. The integration of meta-learning to optimize the discount factor (DF), a vital hyperparameter in RL-based approaches, further enhances the adaptability and convergence rate of the control system, ensuring power quality. Afterwards, the research addresses the challenges in maximum power point tracking (MPPT) for the wind energy conversion systems. Traditional methods like Perturb and Observe (P&O) and incremental conductance are known for their slow dynamic response and susceptibility to steady-state oscillations around the maximum power point, especially under rapidly changing wind conditions. The proposed customized MARL approach overcomes these limitations by employing multiple agents that work collaboratively, resulting in improved energy output and responsiveness to wind speed variations. The use of a meta-learned discount factor optimizes the MARL algorithm, reducing learning duration and enhancing convergence. Extensive simulations and also a 1000W prototype implementation demonstrate the MARL strategy's superiority over traditional MPPT methods, confirming its practical benefits and reliability in real-world applications. Finally, the thesis explores power prediction and management, as well as energy scheduling, in a microgrid (MG) environment. The MG integrates renewable energy sources such as wind turbines (WT), photovoltaic (PV) systems, and battery energy storage systems (BESS), along with combined cooling, heating, and power (CCHP) units. Traditional forecasting methods, such as ARIMA models and simple neural networks, often fail to capture the complex temporal dependencies and variability in renewable energy sources, leading to inaccurate predictions. A multi-layer recurrent neural network (MLRNN) is developed for accurate 24-hour forecasting of renewable energy generation. A grid-search method is proposed to optimally tune the number of RNN layers and the optimizer learning rate. This model leverages historical wind and solar data to capture complex temporal dependencies and patterns. The predicted values are then used as the maximum WT and PV output capacity to optimize power management within the MG using the proposed MARL method. This approach minimizes fuel and CO₂ emissions costs, enhances coordination among MG components, and ensures efficient power distribution, resource utilization, and BESS scheduling. Traditional centralized control methods can be computationally intensive and less responsive to real-time changes, whereas the decentralized control provided by MARL will reduce computational burden and improves response quality, demonstrating its effectiveness in maintaining optimal MG performance. Comparative analysis validates the effectiveness of the approach.

Overall, this thesis provides a robust and innovative framework for enhancing wind energy conversion, renewable energy forecasting, and microgrid power management through the application of a unified MARL-based approach. For the coding and simulation Python programming and Simulink MATLAB are used, respectively. The findings underscore the potential of MARL application to significantly improve the efficiency, reliability, and environmental sustainability of renewable energy systems and WECS