Lossless Data Compression with Bit-back Coding on Massive Smart Meter Data
Heehun Jeong, Giup Seo, Euiseok Hwang · 2022 IEEE International Conference on Big Data (Big Data) · 2022
In this paper, lossless time-series data compression scheme with bit-back asymmetric numeral systems (BB-ANS) is proposed for massive smart meter environment. As smart meters increase in deployment and connectivity, an efficient compression method is needed to transmit and save big data. Bit-back coding was introduced as a novel compression method using bayesian inference modeling. Recently, bit-back coding is combined with asymmetric numeral systems (ANS) which are stack-like structures, showing significant compression gains on several cases. ANS is an approach for entropy coding combining Huffman coding and arithmetic coding which has a first-in-last-out (FILO) form suited for bit-back coding. Compared to other compression methods, bit-back coding effectively shares priorly learned probabilistic models for encoder and decoder. In this study, variational auto-encoder (VAE) is customized to jointly learn approximate posterior and likelihood between message and latent variables. Thus, the smart meter data can be efficiently updated within finite length time-intervals. The proposed scheme is evaluated with the actual smart meter dataset and the results demonstrate the superiority of BB-ANS in compression ratios over other state-of-the-art lossless compression schemes. To the best of our knowledge, this study is the first attempt to apply bit-back coding to time-series smart metering data, enabling efficient data compression with deep generative models.