A Two-Stage Deep Autoencoder-Based Missing Data Imputation Method for Wind Farm SCADA Data

This paper proposes a novel two-stage method for imputing missing SCADA data of wind turbines with high accuracy based on deep nonparametric models, sparse autoencoders (SAE), and a gradient-based optimization algorithm, coordinate descent (CD). A complex pattern of missing data, namely, data loss of correlated attributes (DLCA) that occurs simultaneously, is focused on and studied. In this paper, the missing data imputation is formulated as a two-stage optimization problem. In the first stage, the reconstruction error (RE) of SAE is regarded as the loss for training nonparametric attribute reconstruction models via a complete dataset to learn a low-dimensional manifold, in which data are densely distributed. At the second stage, RE serves as an objective function for optimizing the missing data imputation of a similar but incomplete dataset based on the developed SAE. According to the potential convexity of REs with respect to the imputation of missing attributes, which is empirically discovered through preliminary experiments, the CD algorithm is applied to efficiently solve the optimization problem. The efficacy of the proposed method is validated by using a large real wind turbine dataset. The results of the computational experiments demonstrate that the proposed method performs well on considered benchmarks that are well known for imputing missing data.