Frequency Selective Auto-encoder for Smart Meter Data Compression

Abstract

With the development of the internet of things (IoT), the power grid has become intelligent using massive IoT sensors, such as smart meters. Broadly installed smart meters can collect large amount of data to improve grid visibility and situational awareness. On the other hand, the limited capacities of storage and communication may result in constraints such as an overload of the infrastructure in IoT environment. This increases the needs for efficient and various compression techniques and in response, deep learning-based compression techniques such as an auto-encoder (AE) have recently been deployed. However, existing models may show limited compression performance when the spectral properties of high frequency sampled power data are widely varying over time. In this paper, the AE compression model based on a frequency selection method is proposed to improve the reconstruction quality while maintaining the compression ratio (CR). The proposed method aims to compress data efficiently by selectively applying the customized compression models, depending on the spectral properties of the corresponding time windows. Specifically, a framework of the proposed method includes two parts: (i) the power data is divided into a series of time windows with specified spectral properties, e.g. high-frequency, medium-frequency, or low-frequency dominance, and (ii) the AE models suitable for each frequency characteristics are separately trained and selectively applied for power data compressions. Through simulation using the DRED data set, it was confirmed that the proposed automatically frequency-selective AE model shows significantly improved reconstruction performance compared to the existing model with the same CR. In addition, the proposed model shows the reduction in computational complexity through analysis of the learning process.