An AI-based, Error-bounded Compression Scheme for High-frequency Power Quality Disturbance Data

TL;DR

This work tackles data deluge from high-frequency power-quality monitoring by proposing an AI-based, multi-stage compression scheme that guarantees a deterministic reconstruction error bound. It combines CS and CNN/RNN autoencoders in stage 1, lossy but bounded stage-2 compression (quantization, zfp, SZ3), a residual-bounding stage 3 with masking and differential encoding, and a final lossless stage 4. The approach achieves compression rates from $5$ to $68$ and retains downstream task performance, with classification accuracy degradation between $0.8\%$ and $11.9\%$ depending on the bound and noise. The findings indicate the method can substantially reduce data size while preserving analytical viability, though simpler, non-AI methods may suffice under certain conditions; the authors also highlight robustness to noise and provide publicly available training data for future research.

Abstract

The implementation of modern monitoring systems for power quality disturbances have the potential to generate substantial amounts of data, reaching a point where transmission and storage of high-frequency measurements become impractical. This research paper addresses this challenge by presenting a new, AI-based data compression method. It is based on existing, multi-level compression schemes; however, it uses state-of-the-art technologies, such as autoencoders, to improve the performance. Furthermore, it solves the problem that such algorithms usually cannot ensure an error bound. The scheme is tested on synthetically generated power quality disturbance samples. The evaluation is performed using different metrics such as final compression rate and overhead size. Compression rates between 5 and 68 were achieved depending on the error bound and noise level. Additionally, the impact of the compression on the performance of subsequent algorithms is determined by applying a classification algorithm to the decompressed data. The classification accuracy only declined by 0.8--11.9 \%, depending on the chosen error bound.

Figures (8)

• Figure 1: Proposed multi-stage compression method
• Figure 2: Comparison of compression rates for different algorithm combinations on power quality disturbance data
• Figure 3: Residuals larger than 0.001 after applying CS with a compression rate of 16 and either no further compression, zfp, or quantization
• Figure 4: Comparison of average relative compressed size of data, focusing on the overhead introduces by residuals
• Figure 5: Comparison of the most efficient compression schemes for varying error bounds and noise level