Multi-indicator Water Time Series Imputation with Autoregressive Generative Adversarial Networks

The water quality data has missing values and lacks integrity because water environment monitoring equipments are easily damaged by environmental influences, thereby affecting the analysis accuracy of downstream tasks. Traditional data imputation methods include mea/Mast filling, K-nearest neighbor, matrix factorization, Lahrangian interpolation, etc., do not consider time dependence or fail to use complex relations among multiple features. Inspired by successful applications of various variants of Generative Adversarial Networks (GANs) on time series data, this work proposes a time series data imputation method called GEDA, which integrates -GAN, an Encoder-Decoder structure, and an Autoregressive network. GEDA adopts GAN to learn the probability distribution of multi-feature time series, and imputes the missing values with the generated data. Then, GEDA combines feature extraction of the encoder-decoder structure, and time dependence capturing of the autoregressive network. Real-life dataset-based experimental results demonstrate GEDA outperforms several state-of-the-art data imputation methods in terms of accuracy.