最大功率點追蹤技術常應用於光伏系統中以便能在任何環境下擷取系統的最大功率。擾動與觀察法為最知名的實作方法之一,然而因為擾動步幅被設為固定大小,此方法有可能使最大功率點附近產生大幅震盪或是導致低落的追蹤效率。本論文提出一個基於強化學習的變動步幅最大功率點追蹤法,且使用建立Q表的Q學習演算法來實現。另外,也使用Q網路來實作,並將此視為一個泛化性更佳的方法。和擾動與觀察法相似的是,本方法不需要事先取得關於實際光伏模組的資訊,另外,藉由區分學習階段與追蹤階段,本方法可以實現離線的最大功率點追蹤。擾動與觀察法、使用Q表以及Q網路實作的基於強化學習之最大功率點追蹤法之模擬與實驗結果皆有呈現在本篇論文當中。兩個提出的方法都有較小的漣波以及較快的追蹤速度,故可得知其表現都比擾動與觀察法優良。
Maximum power point tracking technique is often used in photovoltaic (PV) system to extract the maximum power at any environment condition. The perturb and observe (P&O) method is one of the most well-known MPPT methods. However, large oscillations around maximum power point (MPP) or low tracking efficiency may happen since the perturbation is set fixed-sized. In this thesis, a reinforcement learning based variable step size maximum power point tracking (RL MPPT) method is proposed. Q-learning is used as the algorithm of the proposed methods and is implemented by constructing the Q-table (RL-QT MPPT). A Q-network approach (RL-QN MPPT) is also proposed as a more general representation of the RL MPPT method. Similar to the P&O method, implementing of the algorithm doesn’t require the information of the actual PV module in advance. Additionally, by separating the learning phase and the tracking phase, the proposed system is able to track the MPP offline. The simulation and the experiment results of the P&O method, the RL-QT MPPT method, and the RL-QN MPPT method are all presented in this thesis. With smaller ripples and faster tracking speed, the two proposed method outperform the P&O method.