| 摘要: | 隨著全球環保意識逐步提升，綠色能源取代傳統化石燃料發電已成為未來能源發展的重要趨勢。然而以太陽能與風力發電為代表的再生能源具有顯著的間歇性與不穩定特性，使得儲能系統成為維持電網穩定性不可或缺的配套設施。近年來全球天災事件頻繁發生，傳統電網的集中式發電、輸電、配電結構暴露出明顯的脆弱性與韌性不足問題。為此微電網的概念逐漸受到重視，透過在用戶端安裝的分散式能源系統，這種架構不僅降低對傳統輸配電網的依賴，能在主電網發生故障或電力中斷的情況下，持續維持用戶端的穩定供電。
本文進一步導入模組化聚落式微電網的概念，將多個微電網透過聚落的方式進行互聯，以達成能源資源的共享與調度。此架構在發生電力中斷或自然災害時，各個聚落內部或跨聚落間的能源能夠有效互相交易，顯著提升整體聚落微電網系統的韌性與可靠性。另外透過分散式能源共享機制，還能有效提高太陽能發電功率的利用率，達到能源使用的最佳化。
為了更有效地管理與調度這種複雜且動態的系統，本文提出一種基於Principal Component Analyze（PCA）與Twin Delayed Deep Deterministic Policy Gradient（TD3）整合的強化學習方法PCA-TD3；傳統最佳化方法多半僅能對當前時刻的即時狀態進行局部最佳化，無法全面考慮狀態變化，而PCA-TD3透過不斷與環境互動，以最大化長期累積獎勵為目標，這種方法使得控制策略更全面，能有效避免局部最佳解所帶來的控制震盪與不穩定性。
在本文的測試中在併網模式下PCA-TD3相比PSO節省27.3%的運轉成本，孤島模式中兩天的模擬結束後可以比PSO多出4.7%的剩餘能量。綜合而言，本文提出的方法不論是在併網模式或是孤島模式都展現出優秀的控制能力。;With the growing global awareness of environmental protection, replacing tradi-tional fossil fuel power generation with green energy sources has become an essential trend for future energy development. However, renewable energy sources such as solar and wind power have significant intermittency and instability issues, making energy storage systems indispensable for maintaining grid stability. In recent years, frequent natural disasters worldwide have highlighted the vulnerability and lack of resilience inherent in traditional centralized power generation, transmission, and distribution grids. Consequently, the concept of microgrids has gained considerable attention, pro-moting distributed energy systems installed at the consumer level. Such structures not only reduce dependence on conventional transmission and distribution grids but also ensure continuous and stable power supply to consumers during grid failures or power outages.
This study further introduces the concept of modular rundling microgrids, which interconnect multiple microgrids into rundling for effective energy resource sharing and allocation. This architecture significantly enhances the resilience and reliability of the overall rundling microgrid system, particularly during power outages or natural disas-ters. Additionally, the decentralized energy-sharing mechanism effectively improves solar energy utilization, optimizing energy usage.
To efficiently manage and dispatch such complex and dynamic systems, this study proposes a reinforcement learning approach that integrates Principal Component Anal-ysis (PCA) and Twin Delayed Deep Deterministic Policy Gradient (TD3), referred to as PCA-TD3. Traditional optimization methods typically optimize only instantaneous states, failing to account comprehensively for state transitions and future conditions. In contrast, PCA-TD3 continuously interacts with the environment, aiming to maximize long-term cumulative rewards. This comprehensive approach effectively mitigates con-trol oscillations and instability issues caused by local optimum solutions, ensuring more stable and effective management strategies.
Experimental results indicate that, in grid-connected mode, PCA-TD3 achieves a 10.3% reduction in operational costs compared to Particle Swarm Optimization (PSO). In island mode, after a two-day simulation, PCA-TD3 retains 4.7% more remaining energy compared to PSO. Overall, the proposed method demonstrates superior control capabilities in both grid-connected and islanded scenarios, highlighting its effectiveness and practical applicability. |
