以強健混和古典和量子模型進行短期風速預測

A Robust Hybrid Classical and Quantum Model for Short-Term Wind Speed Forecasting

洪絲婷(CHRISTINE JOY ELQUIERO ARCE)

指導教授 ： 洪穎怡

中原大學/電機資訊學院/電機工程學系國際碩士班/碩士(2023年)

https://doi.org/10.6840/cycu202301107

摘要

電是18世紀最偉大的發現之一。然而，儘管這項努力帶來了巨大的好處，但它也帶來了一些問題，包括大量污染和不可再生能源幾近枯竭耗盡。由於這些問題，涉及無限自然資源、幾乎不產生污染且不存在稀缺風險的可再生能源發電越來越受歡迎。整合可再生能源發電（如風力發電）雖然有益，但已使電力公司的電力調度成為更難解決的問題。這是由於風力發電固有的波動性和不可預測性。為了解決這種不確定性，必須開發一種準確的風速預報技術。在本論文中，開發了一種混合經典-量子模型來研究長短期記憶(LSTM)和量子神經網絡(QNN)兩種模型的特性，並利用它們的優勢。LSTM是一種深度學習模型，通常用於時間序列數據，因為它能夠存儲信息並隨時取用它。因此，對於學習數據序列和解決傳統遞迴神經網路造成的梯度消失問題是有效的。同時，QNN的行為類似於改進的機器學習模型，因為它們具有強大的表達能力。一串量子閘的效果取決於它們在QNN中的設計方式。 QNN由利用量子力學原理以實現量子優勢的量子閘組成。此外，穩健設計的田口方法以最優設計獲得最佳結果的，並且對季節性變化的數據不敏感。田口方法強調正交實驗，這些實驗旨在系統地在搜索空間做測試，以避免其耗時費力全域搜索。這些正交實驗的結果決定了所提出模型的參數和超參數的最佳組合。本論文涉及不同國家（台灣、中國、韓國和菲律賓）的七個風電場的歷史風速數據。歷史數據用於預測目標風電場—一個位於台灣附近的福海風場的日前風速值。通過田口方法獲得的優化設計的結果，隨後與現有的 深度學習和機器學習模型進行了比較。比較模擬結果顯示，所提出的穩健混合經典量子模型優於常用傳統模型，如非線性自迴歸網路、隨機森林法、極端梯度提升法、支持向量迴歸和RNN-LSTM。

關鍵字

深度學習 ； 量子神經網絡 ； 穩健設計 ； 風速預測

並列摘要

Electricity is, undeniably, one of the greatest discoveries of the 18th century. However, despite the tremendous advantages that resulted from this endeavor, it has also caused a handful of problems, including vast pollutions and the near-depletion of non-renewable resources. Renewable power generation, which involves unlimited natural resources that yield almost no pollution and do not run the risk of scarcity, gained increasing popularity due to these problems. Albeit beneficial, integrating renewable power generation, like wind power generation, has made power scheduling by power utilities a tougher problem to solve. This is due to the innate volatility and unpredictability of natural resources. To address this uncertainty, an accurate wind speed forecasting technique must be developed. In this thesis, a hybrid classical–quantum model is developed to investigate the characteristics of two models, a long short-term memory (LSTM) and a quantum neural network (QNN), and exploit their advantages. LSTM is a deep learning model generally used for time series data because it is capable of storing information, and accessing it at any time. Thus, LSTM is effective for learning sequences of data and for solving the vanishing gradients problem, which occurs in the traditional recurrent neural network (RNN). Conjointly, QNNs, act like improved machine learning models, as they are characterized by having great expressive power. The effects of a series of quantum gates vary depending on the way they were designed in a QNN. QNNs are made up of quantum gates that exploit the principles of quantum mechanics in order to achieve quantum advantage. Furthermore, the Taguchi method of robust design is used to obtain the optimal design that yields the best results, and is the most insensitive to changes in the seasonal data. The Taguchi method highlights orthogonal experiments that are designed to systematically test in the search space that would otherwise be extremely time-consuming and laborious. The results of these orthogonal experiments dictate the optimal combination of parameters and hyperparameters of the proposed model. This thesis involves historical wind speed data from seven wind farms located in different countries: Taiwan, China, South Korea, and the Philippines. The historical data are used to forecast the 24 hour-ahead wind speed values at the target wind farm – Fuhai, which is located near Taiwan. The results from the optimal design obtained through Taguchi method are consequently compared with existing deep learning and machine learning models. Comparative simulation results show that the proposed robust hybrid classical-quantum model outperforms the popularly used models, such as nonlinear autoregressive network, random forest, extreme gradient boosting, support vector regression, and RNN-LSTM.

並列關鍵字

Deep Learning ； Quantum Neural Network ； Robust Design ； Wind Speed Forecasting

參考文獻

References

[1] Zohuri, B. (2016). Electricity, an Essential Necessity in Our Life. In: Application of Compact Heat Exchangers For Combined Cycle Driven Efficiency In Next Generation Nuclear Power Plants. Springer, Cham. https://doi.org/10.1007/978-3-319-23537-0_2

[2] “Frequently asked questions (faqs) - U.S. energy information administration (EIA),” Frequently Asked Questions (FAQs) - U.S. Energy Information Administration (EIA), 2021. [Online]. Available: https://www.eia.gov/tools/faqs/faq.php?id=38&t=6. [Accessed: 14-Dec-2022].

[3] International Energy Agency, “Global Electricity Demand Set To Slow in 2022 and 2023”, 20 July 2022.

[4] D. Elliott, Sustainable energy: Opportunities and limitations. Basingstoke: Palgrave Macmillan, 2010.