隨著電動運輸的發展，電動巴士充電站的高變動類負載對既有電網造成巨大的壓力，因此結合太陽能與電池儲能系統（Battery Energy Storage System， BESS）的綠色微電網成為具發展潛力的解決方案。然而，傳統情境下的儲能容量評估多未能整合硬體規劃與動態排程控制，且常忽略電池的非線性老化機制，導致評估結果產生明顯偏差。 為了解決此問題，本研究提出一套基於深度強化學習（Deep Reinforcement Learning， DRL）的表後儲能系統容量配置與經濟調度最佳化框架。本研究將時間電價、需量電費限制，以及基於雨流計數法與JEITA模型的電池老化動態，整合於統一的馬可夫決策過程（Markov Decision Process， MDP）中。其控制核心採用連續動作空間的深度確定性策略梯度演算法(Deep Deterministic Policy Gradient， DDPG)，並設計多目標獎勵函數，建立兼顧經濟效益、電池健康度與系統穩定性的連續功率控制策略。 透過十二種天氣與時間電價情境進行模擬，實驗結果顯示DDPG演算法在連續功率控制上顯著優於傳統規則式控制與雙重深度Q網路演算法(Double Deep Q-Network， D-DQN)方法。DDPG能有效最大化電價套利及削峰填谷的經濟效益，同時將電池的老化衰退維持在合理且接近保守規則式控制的範圍內。最終的經濟與投資回收分析指出，在最佳與最差情境下，不同的功率轉換設備容量配置能展現最佳的綜合經濟效益與投資穩健性。本研究為電動巴士充電站微電網提供了具實務可行性的儲能容量規劃與調度依據。;With the rapid development of electric transportation， the highly variable charging demand of electric bus stations imposes significant stress on existing power grids. In response to these challenges， green microgrids integrating photovoltaic generation and battery energy storage systems have emerged as a promising and sustainable solution. However， conventional BESS capacity planning methods often decouple hardware sizing from dynamic operational control and frequently overlook the nonlinear battery degradation mechanisms， leading to considerable deviations in economic and operational assessments. To address these limitations， this study proposes a deep reinforcement learning based framework for optimal capacity sizing and economic dispatch of behind-the-meter battery energy storage systems. Time-of-use electricity pricing， demand charge constraints， and battery degradation dynamics modeled through the Rainflow Counting Method and the JEITA degradation model are integrated into a unified Markov Decision Process . The control strategy is developed using the Deep Deterministic Policy Gradient algorithm， which is well suited for continuous action spaces. Furthermore， a multi-objective reward function is designed to simultaneously consider economic performance， battery health preservation， and system operational stability. Simulation studies are conducted under twelve scenarios combining different weather conditions and electricity tariff structures. The results demonstrate that the proposed DDPG-based approach significantly outperforms conventional Rule-Based Method control and Double Deep Q-Network methods in continuous power control applications. The DDPG agent effectively maximizes economic benefits through energy arbitrage and peak-shaving operations while maintaining battery degradation at a reasonable level comparable to that achieved by conservative Rule-Based Method strategies. Finally， economic and payback analyses indicate that different power conversion system capacity configurations exhibit optimal economic performance and investment robustness under both best-case and worst-case operating conditions. The proposed framework provides a practical and effective methodology for BESS capacity planning and operational scheduling in electric bus station microgrids.
