改良式人工智慧方法於電力系統之預測應用;Improved Artificial Intelligence-Based Methods for Power System Forecasting Applications

[[abstract]]近年來環保意識逐漸提升，許多國家已逐漸意識到開發運用再生能源之必要性，目前已有越來越多的國家積極投入再生能源的開發工作，其中又以取之不盡的風力能源、太陽能源之相關研究最受重視。由於隨著風力發電與太陽光電系統的建置與容量逐漸擴大，再生能源預測技術對於減輕太陽光電與風力發電之輸出的隨機性影響具有相當重要的意義，準確的再生能源發電預測則可提高電網穩定性、降低備載容量、以及電力系統運轉成本及減少污染排放，本論文針對風力發電廠與太陽光電廠之發電輸出值進行不同資料解析度之預測。然而於大型負載設備（例如：風力發電機組或電弧爐）運轉後所產生有關電力品質之電壓閃爍問題也值得探討，若這些不良的電壓閃爍嚴重程度也可以被預測，則電力公司與大型負載機組設備用戶之間可以共同利用改善設備（例如：靜態虛功補償器或熔爐控制）進行改善電壓閃爍問題。因此，本論文採用煉鋼廠的電弧爐設備為研究對象預測三種電壓閃爍指標（ΔV10, Pst, and Plt）之嚴重程度。本論文即依據上述之問題，將提出改良式輻狀基底類神經網絡（IRBFNN）、具有誤差回授機制之輻狀基底類神經網絡（RBFNN-EF）、具有誤差回授機制之改良式輻狀基底類神經網絡（IRBFNN-EF）、高斯混合模型之類神經網絡（GMMNN）、結合灰色理論之深度學習類神經網絡（Grey-DLNN）、結合灰色理論之輻狀基底類神經網絡（Grey-RBFNN）、結合灰色理論之改良式輻狀基底類神經網絡（Grey-IRBFNN）與結合查表法之灰色模型（Grey-LUT）進行風速、風力、太陽光以及電壓閃爍嚴重程度之預測應用並與過去傳統之類神經網絡方法進行比對，測試結果顯示，本論文所應用之預測方法於實際電力系統之可?性。 Increasing renewable energy penetration has been a global trend in the past decades. Wind and solar power are expected to contribute significantly to the renewable energy targets owing to advancements in renewable energy technologies, abundance of free resource and commercial viability. Therefore, the predictability of wind and solar power in managing load and generation balance is crucial to system operations. This thesis proposes improved models for wind and solar power output at different interval forecast. Also, due to the operation of the large electric devices (e.g. wind power generator, electric arc furnace, etc.) which would make the power system produce serious voltage flicker. If the flicker levels are predictable, corrective solution such as static var compensation may be developed for both electric utilities and the customer. This thesis adopts electric arc furnace at steel industrial company as investigated target for three indices of flicker severity (ΔV10, Pst, and Plt) forecast.This thesis proposes improved artificial intelligence methods to solve above-mentioned problems which include the improved radial basis function neural network (IRBFNN), radial basis function neural network with an error feedback (RBFNN-EF), improved radial basis function neural network with an error feedback (IRBFNN-EF), Gaussian mixture model neural network (GMMNN), radial basis function neural network integrated with grey theory (Grey-RBFNN), improved radial basis function neural network integrated with grey theory (Grey-IRBFNN), deep learning neural networks integrated with grey theory (Grey-DLNN), and Grey theory with look-up table (Grey-LUT). Performance comparisons between the proposed and traditional methods are reported for wind speed at 10-minute interval, wind power at 10-minute interval, wind power output at one-minute interval, solar power generation at one-minute interval, and three kinds of flicker severity level forecast. Simulated results (i.e. wind speed and power forecast, solar power forecast, and flicker severity forecast) can provide more accurate and effective forecast than other compared methods to the actual power system forecasting problems