There has been a shift towards energy sustainability in recent years, and this shift should continue. The steady growth of energy demand because of population growth, as well as heightened worries about the number of anthropogenic gases released into the atmosphere and deployment of advanced grid technologies, has spurred the penetration of renewable energy resources (RERs) at different locations and scales in the power grid. As a result, the energy system is moving away from the centralized paradigm of large, controllable power plants and toward a decentralized network based on renewables. Microgrids, either grid-connected or islanded, provide a key solution for integrating RERs, load demand flexibility, and energy storage systems within this framework. Nonetheless, renewable energy resources, such as solar and wind energy, can be extremely stochastic as they are weather dependent. These resources coupled with load demand uncertainties lead to random variations on both the generation and load sides, thus challenging optimal energy management. This thesis develops an optimal energy management system (EMS) for a grid-tied solar PV-battery microgrid. The goal of the EMS is to obtain the minimum operational costs (cost of power exchange with the utility and battery wear cost) while still considering network constraints, which ensure grid violations are avoided. A reinforcement learning (RL) approach is proposed to minimize the operational cost of the microgrid under this stochastic setting. RL is a reward-motivated optimization technique derived from how animals learn to optimize their behaviour in new environments. Unlike other conventional model-based optimization approaches, RL doesn't need an explicit model of the optimization system to get optimal solutions. The EMS is modelled as a Markov Decision Process (MDP) to achieve optimality considering the state, action, and reward function. The feasibility of two RL algorithms, namely, conventional Q-learning algorithm and deep Q network algorithm, are developed, and their efficacy in performing optimal energy management for the designed system is evaluated in this thesis. First, the energy management problem is expressed as a sequential decision-making process, after which two algorithms, trading, and non-trading algorithm, are developed. In the trading algorithm case, excess microgrid's energy can be sold back to the utility to increase revenue, while in the latter case constraining rules are embedded in the designed EMS to ensure that no excess energy is sold back to the utility. Then a Q-learning algorithm is developed to minimize the operational cost of the microgrid under unknown future information. Finally, to evaluate the performance of the proposed EMS, a comparison study between a trading case EMS model and a non-trading case is performed using a typical commercial load curve and PV generation profile over a 24- hour horizon. Numerical simulation results indicated that the algorithm learned to select an optimized energy schedule that minimizes energy cost (cost of power purchased from the utility based on the time-varying tariff and battery wear cost) in both summer and winter case studies. However, comparing the non-trading EMS to the trading EMS model operational costs, the latter one decreased cost by 4.033% in the summer season and 2.199% in the winter season. Secondly, a deep Q network (DQN) method that uses recent learning algorithm enhancements, including experience replay and target network, is developed to learn the system uncertainties, including load demand, grid prices and volatile power supply from the renewables solve the optimal energy management problem. Unlike the Q-learning method, which updates the Q-function using a lookup table (which limits its scalability and overall performance in stochastic optimization), the DQN method uses a deep neural network that approximates the Q- function via statistical regression. The performance of the proposed method is evaluated with differently fluctuating load profiles, i.e., slow, medium, and fast. Simulation results substantiated the efficacy of the proposed method as the algorithm was established to learn from experience to raise the battery state of charge and optimally shift loads from a one-time instance, thus supporting the utility grid in reducing aggregate peak load. Furthermore, the performance of the proposed DQN approach was compared to the conventional Q-learning algorithm in terms of achieving a minimum global cost. Simulation results showed that the DQN algorithm outperformed the conventional Q-learning approach, reducing system operational costs by 15%, 24%, and 26% for the slow, medium, and fast fluctuating load profiles in the studied cases.
Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:uct/oai:localhost:11427/37595 |
Date | 31 March 2023 |
Creators | Muriithi, Grace |
Contributors | Chowdhury, Sunetra |
Publisher | Faculty of Engineering and the Built Environment, Department of Electrical Engineering |
Source Sets | South African National ETD Portal |
Language | English |
Detected Language | English |
Type | Master Thesis, Masters, MSc |
Format | application/pdf |
Page generated in 0.0012 seconds