應用深度學習即時評估極端風況下的風機負載

Abstract

台灣海峽在秋冬季節有穩定的東北季風，而夏季則受到西南氣流影響，加上狹窄的地形效應，台灣在開發離岸風能方面具有很大的潛力。然而，由於台灣緊鄰太平洋且位於中緯度地區，使得台灣容易受到颱風侵擾，颱風所造成的極端負載往往是風機結構損壞的主因。因此，如何在颱風侵襲下即時評估風機的極端負載顯得至關重要。 本研究使用中尺度天氣研究預報模式(Weather Research and Forecasting, WRF)進行歷年颱風的風場模擬，並結合風機尺度隨機全場紊流模擬器Turbsim，模擬出目標點之颱風下風機尺度紊流風場。將此紊流風場輸入至OpenFAST的入流風模組中，以模擬颱風造成風機的極端負載。為了達成颱風侵襲時即時評估風機的極端負載，將WRF模擬的颱風相關數據加上鄰近目標點之氣象測站數據，訓練兩種深度學習模型，分別為深度神經網路(deep neural networks, DNN)及長短期記憶(long short-term memory, LSTM)，以計算目標點風速及風機極端負載。 將OpenFAST模擬之極端負載結合WRF模擬風場進行觀察與分析，以瞭解颱風侵襲的極端負載變化特性及類型。本研究依照極端負載變化特性將颱風分為三種，分別為偏向颱風、穿越颱風、低風速颱風。偏向颱風及穿越颱風的取名方式是根據颱風中心經過目標點的方式，偏向颱風為颱風中心無經過觀測塔，僅暴風圈碰觸到觀測塔，而穿越颱風為颱風中心經過觀測塔上方。此兩種颱風造成目標點風速達到切出風速(cut-out wind speed)，導致風機停機負載突降，以及風速低於切出風速，風機重新運轉負載突增等的負載變化特性。低風速颱風造成目標點風速未達到切出風速，無負載變化特性呈現，而有此名稱。藉由此觀察之特性、趨勢以評估DNN及LSTM兩種模型計算極端負載的效果。 訓練完成的兩種模型用以計算目標點之風速，並與WRF模擬結果及實際量測資料進行比對、驗證，結果顯示三種模型計算結果與量測值的趨勢相符，WRF模擬結果表現較佳，而深度學習模型中LSTM表現優於DNN。將DNN及LSTM模型進行颱風侵襲下風機負載評估，與OpenFAST模擬結果比較負載特性及趨勢，結果顯示DNN及LSTM模型計算結果與OpenFAST模擬結果趨勢相近，但DNN 無法表現穿越颱風的負載特性，LSTM在特性表現大致上較DNN佳。另外，若將僅有負載特性的偏向颱風及穿越颱風作為數據集給DNN及LSTM模型訓練，也無法改善負載特性的計算效果，原始數據訓練的LSTM依然表現較佳。 總結，本研究建立一完整的模擬流程，得到能在颱風侵襲時即時評估風機極端負載的深度學習模型，且在風速計算結果中，與實際量測值比較後顯示模型具有一定的準確度及可行性，期許本研究之模擬流程及模型，能實際應用於颱風侵襲時即時評估風機極端負載。

The Taiwan Strait experiences stable northeast monsoon during the autumn and winter seasons, while in summer it is influenced by southwesterly flow. Coupled with the narrow terrain effect, Taiwan holds significant potential for developing offshore wind energy. However, due to Taiwan's proximity to the Pacific Ocean and its location in the mid-latitude region, it is frequently affected by typhoons. The extreme loads caused by typhoons are often the main reason for structural damage to wind turbines. Consequently, it is of paramount importance to conduct real-time assessments of the extreme loads on wind turbines during typhoon invasions. In this study, the Weather Research and Forecasting (WRF) model was used to simulate the wind fields of historical typhoons. In addition, the full-field stochastic turbulence simulator Turbsim was used to simulate the turbulent wind field at the target point under typhoon conditions. This turbulent wind field was then fed into the OpenFAST inflow wind module to simulate the extreme loads on wind turbines caused by typhoons. In order to achieve a real-time assessment of the extreme loads on wind turbines during typhoon events, the typhoon-related data from the WRF simulations, together with meteorological station data near the target point, were used to train two deep learning models: deep neural networks (DNN) and long short-term memory (LSTM). These models were used to calculate the wind speeds at the target point and the extreme loads on the wind turbines. By integrating the extreme load simulations from OpenFAST with wind fields simulated by WRF, this study aims to observe and analyze the characteristics and types of extreme loads during typhoon events. Based on the characteristics of extreme load variations, typhoons are classified into three categories: deflected typhoons, crossing typhoons, and low wind speed typhoons. The names of deflected and crossing typhoons are based on the wind direction at the target point as the typhoon center passes. Each of these two types of typhoons has different load variation characteristics. In contrast, low wind speed typhoons do not cause the wind speed at the target point to reach the cut-off wind speed, resulting in no load variation characteristics. By observing these characteristics and trends, the effectiveness of using DNN and LSTM models to calculate extreme loads is evaluated. The trained models, DNN and LSTM, were used to calculate the wind speed at the target point and compared with WRF simulation results and actual measurement data for validation. The results showed that the trends of the calculated values from all three models matched the measured values, with the WRF simulation performing best and the LSTM outperforming the DNN. The DNN and LSTM models were then used to evaluate wind turbine loads under typhoon conditions, and their load characteristics and trends were compared with the OpenFAST simulation results. The results showed that while the load trends calculated by the DNN and LSTM models were similar to those from OpenFAST simulations, the DNN failed to capture the load characteristics of crossing typhoons, while the LSTM generally performed better in characterizing these characteristics. Furthermore, even when only the deflected and crossing typhoons, which have load characteristics, were used as the data set for training the DNN and LSTM models, the calculation of the load characteristics did not improve. The LSTM trained on the original dataset still showed the best performance. In conclusion, this study establishes a comprehensive simulation process and develops deep learning models capable of real-time assessment of wind turbine extreme loads during typhoon invasions. A comparison of the wind speed calculations with actual measurements indicates that the models have a certain degree of accuracy and feasibility. It is anticipated that the simulation process and models developed in this study will be practically applied for real-time assessment of wind turbine extreme loads during typhoon invasions.