Multi-head self-attentive GRU-based missing data repair method for wind turbines

Due to temporary anomalies in sensor readings and additional issues, the data collected by wind turbine monitoring and control systems often feature a considerable number of missing values. This can have detrimental impacts on various critical aspects such as wind power prediction, the calculation of wind power curves, and the implementation of necessary corrections. Traditional approaches to missing data repair in wind turbine datasets rely mainly on machine learning and deep learning techniques. However, methods based on traditional machine learning are often limited by their shallow architectures, which restricts their ability to deeply explore and understand complex data features. On the other hand, while deep learning methods can capture intricate patterns in data, they frequently overlook the temporal dependencies inherent in time-series data. To address these significant challenges and effectively uncover the temporal correlations among the dataset's features, this paper proposes a novel approach for repairing missing data in wind turbines utilizing a multi-head self-attention mechanism within a Gated Recurrent Unit (GRU) framework. The proposed methodology begins with normalizing the dataset and subsequently dividing it into training and testing subsets. Key feature parameters are selected based on the Pearson correlation coefficient to ensure that only the most relevant data is utilized. The core of the approach lies in constructing a GRU model augmented with a multi-head self-attention mechanism, allowing for the restoration of data with proportionate and continuous missing values. Results obtained from the experiments demonstrate that the proposed method not only effectively fills in the missing data but also adequately learns the variation patterns of feature parameters, outperforming traditional GRU and Long Short-Term Memory (LSTM) algorithms in accurately repairing multi-dimensional parameter data.