基於遷移學習於功率預測之風機故障預警

Abstract

本研究使用深度神經網路(Deep Neural Network, DNN)建立風機功率曲線模型(Wind Turbine Power Curve Model)及風機功率預測模型(Wind Turbine Power Prediction Model)，並以遷移學習(Transfer Learning)的訓練方式來提升模型之學習成效。 本研究使用台電麥寮風場15部風機的SCADA(Supervisory Control And Data Acquisition system)資料進行分析，各部風機以過去3個月的資料訓練出對應的風機功率曲線模型，以功率曲線最貼近保證功率曲線(Guaranteed Power Curve, GPC)的風機當選為當月的模範風機，以模範風機前三個月的資料訓練出作為標準之功率預測模型，並根據該模型之訓練結果求出次月用來判斷資料異常與否的閾值(Threshold)，除此之外，將該模型作為預訓練模型(Pre-Training Model)，複製出15個相同的模型對應到各部風機，各自使用前三個月的資料去對自己的模型由遷移學習進行微調(Fine-Tune)，即可得出次月各部風機的功率預測模型。 使用風機次月的功率預測模型，得出每筆資料的預測值與實際值計算殘差(Residual)，計算每日的平均殘差，若平均殘差超過設定之閾值便提出故障預警(Early Warning)。將異常風機的特徵資料分別與其他風機相比，可將預警分為主要感測器異常導致、相關元件異常導致和其他無紀錄元件異常導致等三類，根據此結果提早進行檢修，不但能縮短風機故障發電效能低落的時間，也可避免風機因突發性故障導致的額外毀損，進而提升風力發電之經濟收益。

The algorithm of the deep neural network was used to establish wind turbine power curve model and wind turbine power prediction model in this study. The transfer learning algorithm was used to improve the accuracy of the power prediction model. The SCADA datasets of 15 wind turbines from TaiPower Mai Liao wind farm were analyzed in this study. The history data in the past 3 months of each wind turbine were used to train its own power curve model. An exemplary wind turbine was selected by the minimum root mean square error between the trained power curve model and the guaranteed power curve. The exemplary wind turbine is the best performance wind turbine in the trained period. The standard power prediction model is trained by the history data in past 3 months of the exemplary wind turbine. The trained results were used to calculate the threshold of the anomaly data in the following month. In addition, the exemplary wind turbine power prediction model was used as a pre-training model. The pre-training model was used by other wind turbines and was fine-tuned by the transfer learning algorithm by and their own history data in the past 3 months. Then, the power prediction model of each wind turbine for the following month could be obtained. The trained power prediction model of each wind turbine was used to predict the power. The residual between the actual power and the predicted power could be calculated and the day average residual could also be calculated. An early warning will be issued when the day average residual exceeds the threshold. There are three types of early warning which include the main sensor abnormality, the related component abnormality and the unrecorded component abnormality based on the features of the wind turbines. According to this early warning, the maintenance work can be carried out in advance. It will not only reduce the time of low power generated but also avoid the unexpected damage due to failure of the wind turbine. The early warning model can enhance the economic benefits of wind power generation.