Learning-based Similarity Join for Power Data

Abstract

The increasing instrumentation of physical and computing processes has given us unprecedented capabilities to collect massive volumes of time series. Power data is a typical kind of time series. Considering that the original time series data has ineluctable limitations such as uneven distribution, non-uniform length, poor sampling rate and noisy, we propose a learning=based similarity join for power data consisting of RNN encoder and matrix model. In addition, we develop the partition techniques by grouping process nodes following the matrix join model, ensuring the accuracy and efficiency of similarity join for data series. We conduct experiments on real data-set to evaluate the performance of our approach, demonstrating the effectiveness and scalability of our method.