Spatiotemporal Prediction Based on Feature Classification for Multivariate Floating-Point Time Series Lossy Compression

Abstract

A large amount of time series is produced because of the frequent use of IoT devices and sensors. Time series compression is widely adopted to reduce storage overhead and transport costs. At present, most state-of-the-art approaches focus on univariate time series. Therefore, the task of compressing multivariate time series (MTS) is still an important but challenging problem. Traditional MTS compression methods treat each variable individually, ignoring the correlations across variables. This paper proposes a novel MTS prediction method, which can be applied to compress MTS to achieve a higher compression ratio. The method can extract the spatial and temporal correlation across multiple variables, achieving a more accurate prediction and improving the lossy compression performance of MTS based on the prediction-quantization-entropy framework. We use a convolutional neural network (CNN) to extract the temporal features of all variables within the window length. Then the features generated by CNN are transformed, and the image classification algorithm extracts the spatial features of the transformed data. Predictions are made according to spatiotemporal characteristics. To enhance the robustness of our model, we integrate the AR autoregressive linear model in parallel with the proposed network. Experimental results demonstrate that our work can improve the prediction accuracy of MTS and the MTS compression performance in most cases.