運用集成深度神經網路於太陽能發電量之預測

Photovoltaic Generation Prediction based on Ensemble Deep Neural Networks

彭韻軒(Yun-Hsuan Peng)

指導教授 ： 蘇朝墩

國立清華大學/工學院/工業工程與工程管理學系/碩士(2021年)

摘要

電力在現今人類生活中，已是不可或缺的一環，目前的電能大多是仰賴石化燃料或核分裂過程進行能量轉換所產生，然而這些資源蘊藏量相當有限且分佈不均，在能源轉換的過程中，更會產生具有放射性的分裂產物，進而產生潛在性的環境汙染問題。有鑑於上述能源短缺以及環保相關的議題，許多國家也嘗試推行多種能源轉型，其中透過太陽能、風力發電等再生能源最為興盛，這些方式能有效降低碳足跡、二氧化碳，而目前台灣也積極地推廣設置地面型及屋頂型太陽能發電系統。
欲快速且準確地預測太陽能即時發電量，本研究探討使用深度神經網路的模型所帶來的潛在優勢，並提出一個三階段的發電量預測程序，包括資料前處理，接著使用深度神經網路(Deep Neural Networks)結合集成學習(Ensemble Learning)演算法，訓練出適合的配置模型，最終與多種機器學習模型比較並進行模型評估與最佳模型選取。經由一個真實個案之分析，本研究收集太陽能廠歷史發電量之數據進行預測。執行結果顯示最佳模型為基於深度神經網路的集成式學習，使用此模型能夠使太陽能發電量之預測準確率高達95%，也就是誤差低於10瓩。本研究聚焦於關鍵特徵之數據變化，快速且準確預測及時應有之發電量水準，以提升太陽能廠對於整體發電效能及穩定度之掌控。

關鍵字

太陽能發電 ； 發電量預測 ； 深度神經網路 ； 集成式學習

並列摘要

Electricity plays an indispensable role in human life today. Most of the current electricity must be relied on fossil or nuclear fuel to generate energy. However, these resources are limited and unevenly distributed on the earth. In the process of energy conversion, radioactive fission products will be generated, which will cause potential environmental pollution problems. In view of the above-mentioned issues, many countries are pursing energy transformation, hoping to reduce their carbon footprint and carbon dioxide through renewable energy such as solar and wind power. Currently, Taiwan is also actively promoting the installation of ground-type and rooftop-type solar power systems.
To quickly and accurately predict solar real-time photovoltaic (PV), this study proposes a three-stage procedure that includes data preprocessing; uses ensemble learning based on deep neural networks to train a suitable configuration model. Finally, it compares with multiple machine learning models and conducts model evaluation to select optimal model for power generation forecasting. Through the analysis of a real case, this study collects data on the historical photovoltaic from a solar power plant for prediction. The execution result shows that the most appropriate model is using ensemble learning based on deep neural networks, which can make the accuracy rate up to 95%, that is let the error of prediction lower than 10kW. When entering the key variables of PV system in this predicted model, it will output the timely estimated value of power generation precisely. Therefore, it can enhance the control of the overall efficiency and stability to solar power plant.

並列關鍵字

Photovoltaic (PV) system ； Prediction of generated electric energy ； Deep Neural Network ； Ensemble Learning

參考文獻

Altay, E., & Satman, M. H., Stock market forecasting: artificial neural network and linear regression comparison in an emerging market, Journal of Financial Management & Analysis, Vol. 18, No. 2, pp. 18 (2005).

Alzahrani, A., Shamsi, P., Dagli, C., & Ferdowsi, M., Solar irradiance forecasting using deep neural networks, Procedia Computer Science, Vol. 114, pp. 304-313 (2017).

Awad, M., & Khanna, R., Support vector regression, Efficient learning machines, pp. 67-80 (2015).

Badhiye, S. S., Sambhe, N. U., & Chatur, P. N., KNN technique for analysis and prediction of temperature and humidity data, International Journal of Computer Applications, Vol. 61, No. 14 (2013).

Benhmed, K., Touati, F., Al-Hitmi, M., Chowdhury, N. A., Gonzales, A. S., Qiblawey, Y., & Benammar, M., PV power prediction in Qatar based on machine learning approach, 2018 6th International Renewable and Sustainable Energy Conference (IRSEC), pp. 1-4 (2018).