Cascaded Denoising Convolutional Auto-Encoders for Automatic Recovery of Missing Time Series Data

Abstract

This paper proposes a kind of supervised cascaded denoising convolutional auto-encoders (CDCAE), aiming to accurately recover the missing load data in electric power system. The one-dimensional load data are reshaped as two-dimensional image for data enhancement, which enables the convolutional neural network (CNN) to understand the semantics of load data. Numerical results in comparison with similar day filling (SDF) clearly validate the effectiveness of the proposed CDCAE in accuracy.