Advanced Machine Learning Enabled Modern Power System

The intermittent nature of Photovoltaic (PV) systems requires advanced prediction and management strategies to be incorporated into modern power systems. This thesis investigates a comprehensive Advanced Machine Learning (AML) technique that encompasses short-term power forecasting, energy system optimization for intelligent control, and decentralized energy trading among numerous Energy Communities (ECs). Firstly, an accurate PV nowcasting architecture is developed by integrating Recurrent Neural Networks (RNNs) and Stacked Ensemble Learning (SEL) techniques with the time-varying segmentation of weather observations. Utilizing over six years of data collected from an 81 kW PV system located in the Savona Campus, University of Genova, Italy. The proposed algorithm exhibits enhanced accuracy in one-hour-ahead prediction, as a SEL-based model surpasses multiple Long Short-Term Memory (LSTM) RNNs, and the integration of Seasonal Trend Decomposition (STL) further minimizes the forecasting error. In addition, structurally tuned LSTM RNN models combined with STL achieve superior accuracy compared to the traditional LSTM model. A SEL technique enhances solar output under severe air contamination (soiling) to address the challenges of the metropolitan environment. The SEL model achieves the lowest possible error for panel installations across different orientations, despite the presence of a soiling layer, and maintains good predictive accuracy. This method calculates a soiling index to enhance maintenance and determine the optimal cleaning strategy for solar panels. Furthermore, we integrate accurate forecasting into the campus Energy Management System (EMS) using a receding horizon approach. The EMS system updates its optimization hourly, using dynamic 24-hour forecasting, enabling proactive resource use and effective management of the MG. In the case study, the proposed methods achieve higher annual profitability and reliability than the baseline model in day-ahead scheduling by integrating short-term forecasting from multiple LSTM-based algorithms into the control loop. Beyond the local scheduling, the thesis discusses the role of EC in delivering decentralized flexibility and localized scheduling. The Deep Reinforcement Learning (DRL) based approach simulates and encourages the exchange of end-user flexibility by reshaping adjustable loads and battery storage. The proposed approach aims to avoid sharing flexibility during critical hours and preserve end-user privacy by exchanging flexibility at the aggregator level. It focuses on developing a Proximal Policy Optimization DRL (PPO-DRL)-based load management aggregation from the end-user's perspective. It offers a solution to meet grid demand by adhering to limit constraints and increasing economic benefits for the end-user. The results are validated against the fixed-pricing mechanism for real-time EV charging management, with a minimum Mean Absolute Error (MAE) of 0.0560 W. The finding indicates that the intelligent system efficiently coordinates resources, leverages demand flexibility, and increases communities' willingness to balance supply and demand. Finally, a comprehensive literature review suggests that AML methods play a crucial role as key accelerators for decentralized control in EC. It is essential to note that Reinforcement Learning (RL) techniques provide considerable support in decision-making and control tasks. In contrast, unsupervised and supervised learning surpass them in forecasting and data analysis. The suggested optimization, trading, and forecasting techniques jointly reduce operational uncertainty and overall costs by enhancing reliability. The study demonstrates that holistic AML strategies can significantly improve the stability and economic sustainability of EC abundant in Renewable Energy Sources (RES) by proactively mitigating PV intermittency and enabling prosumers' active participation in flexibility sharing and grid services. Hence, the result highlights a progressive strategy in which accurate PV forecasting and intelligent EMS control reduce energy consumption and enhance system resilience, thereby further advancing a sustainable, decentralized energy system.