| 摘要: | 本研究聚焦於電動車大規模接入所引發的電網負載壓力與排程調控問題。隨著電動車滲透率快速上升，傳統配電網面臨瞬時高功率需求增加、尖峰負載過載與儲能協調失衡等風險，進一步威脅系統穩定性與用戶充電體驗。為此，本研究提出一套基於 Lyapunov Drift-Plus-Penalty 模型之動態平衡式電動車充電排程方法，透過構建兩條虛擬列隊系統壓力列隊 Q(t) 與用戶需求列隊 D(t)，即時反映電網與使用者間的負載需求動態。相較於傳統 Mixed Integer Linear Programming(MILP)、強化學習方法 (Deep Q-Network(DQN), Categorical DQN(C51), Independent Deep Q-Network with LSTM(IDQN+LSTM)) 和 Genetic Algorithm(GA) 等方法，基於 Lyapunov Drift-PlusPenalty 的動態平衡式控制架構不依賴精準預測，且不需要歷史資料支持，就可以快速求解近似最適功率分配策略。
本研究整合再生能源輸入、儲能系統、動態價格模型與用戶參與度權重，透過用戶參與權重計算優先級，使能源分配更有效率，實現多源功率協調與即時排程控制。模擬結果顯示，相較 MILP 、DQN、C51、IDQN+LSTM 和 GA 等方法，本研究所提策略在總購電成本、場站平均積壓時間、平均離站缺口與超過饋線門檻的時數等多項指標上皆展現良好平衡，並具備良好的運算效率與實作彈性。進一步引入客戶優先排序與效用，可以增強充電站資源分配的同時，以客戶效用函數量化使用者對充電之效益與成本感受，使參與程度可以更精準的影響最終功率分配。
;This study focuses on the grid load stress and scheduling challenges arising from the large-scale integration of electric vehicles (EVs). As EV penetration continues to grow rapidly, traditional distribution networks face increasing risks such as sudden surges in power demand, peak-load overloading, and imbalance in coordination with energy storage systems. These issues further threaten system stability and degrade user charging experiences. To address these challenges, this study proposes a dynamic equilibrium smart charging scheduling framework, which constructs two virtual queues the system-pressure queue Q(t) and the user-demand queue D(t) to capture real-time dynamics between grid load conditions and user charging requirements.
Compared with conventional Mixed Integer Linear Programming (MILP), reinforcement learning approaches (Deep Q-Network(DQN), Categorical DQN(C51), Independent
Deep Q-Network with LSTM(IDQN+LSTM)), and Genetic Algorithms (GA), the proposed Lyapunov Drift-Plus-Penalty control architecture does not rely on accurate forecasting or historical data. Instead, it can rapidly compute near-optimal power allocation decisions through an efficient real-time balancing mechanism.
This study integrates renewable energy inputs, an energy storage system, a dynamic pricing model, and user participation weighting. By incorporating user-specific priority weights, the system achieves more efficient energy allocation and realizes multi-source power coordination and real-time scheduling control. Simulation results demonstrate that, compared with MILP, DQN, C51, IDQN+LSTM, and GA, the proposed approach achieves strong overall performance across multiple metrics, including total electricity purchase cost, average station waiting time, average departure deficit, and the number of hours exceeding feeder constraints, while also offering superior computational efficiency and implementation flexibility. Furthermore, by incorporating customer prioritization and utility modeling, the system can quantify user-perceived benefits and costs through a customer utility function, enabling participation levels to more accurately influence final power allocation outcomes. |
