Univariate Time Series Data Correction Method using LSTM Autoencoder with Temporal Distance

High-quality data is essential to increase the reliability of machine learning-based prediction models. For time series data, anomalies significantly reduce the accuracy of prediction models. In this paper, we propose a novel time series data correction method that converts abnormal values of univariate time series data into normal ones. For anomaly detection and correction, we utilize the LSTM Autoencoder model, where we propose a new weight function that considers temporal distance. Through experiments using the open NAB data, we show that our proposed method is superior to the recent conventional methods.