Comparative Analysis of Modern, AI-based Data Compression on Power Quality Disturbance Data

Current day power grids are continuously evolving away from large, synchronously-coupled, centralized generation units towards more decentralized structures. This change has a huge impact on grid inertia, stability, and through that power quality. Monitoring the quality of our power supply systems therefore becomes increasingly important in order to analyze and predict critical grid conditions. However, such monitoring systems can create huge amounts of data, which at some point become infeasible to transmit and store. Therefore, this contribution contains the analysis of two modern, AI-based data compression methods: recurrent autoencoders and convolutional autoencoders. They are compared with the more conventional technique of Compressed Sensing. All methods are applied to a diverse set of mathematically-modeled power quality disturbance data samples. Hyperparameter tuning is performed for the AI-based approaches, which shows the impact of different parameters on the networks’ performance and identifies optimal choices for this power quality use case. It becomes clear, that the recurrent autoencoder’s performance in not en-par with the other two approaches. The Compressed Sensing method introduces only very little compression error on harmonic disturbances, especially for lower compression rates. However, the convolutional autoencoder turns out to deliver a better overall performance.