連續型後設模型求解模擬器參數之應用

Abstract

近年來，由於永續發展及淨零碳排的目標，海洋能源相關議題逐漸獲得重視，洋流能更是當中極具潛力的發電形式。然而，受限於水下渦輪機的建置成本及資源限制，物理實驗難以頻繁及重複進行。本研究設計了不同的抽樣方法進行電腦模擬，使用人工神經網路(artificial neural network, ANN)建構後設模型(surrogate model)，捕捉模擬器的輸入輸出關係，並採用啟發式演算法(metaheuristic)搜尋最佳參數。在本研究的數值分析案例中，後設模型與電腦模擬在測試集的平均百分比誤差小於5%，最佳參數模擬結果與物理實驗的誤差可達到小於14%。

In recent years, marine energy issues have increasingly garnered attentions due to the goal of sustainable development and net-zero carbon emissions, with ocean current energy emerging as a highly promising form. Physical experiments, constrained by the high implementation cost and limited resources, are challenging to conduct frequently and repetitively. This study designs various sampling methods to perform computer simulations and applies artificial neural networks (ANN) to build a surrogate model, capturing the input-output relationships of the simulator. Additionally, we employ metaheuristic algorithms to search for optimal parameters of the simulator. In the investigated numerical examples, the average percentage error between the surrogate model and computer simulation on the test set is less than 5%. The error between the simulation results with optimal parameters and the physical experiments can reach less than 14%.