Intra and inter-series pattern representations fusion network for multiple time series forecasting

Abstract

Multiple time series (MTS) can comprise data collected from various wireless sensor networks in the actual application, and each source provides a distinctive pattern. Most existing neural network methods attempt to model the patterns of individual time series by training a global model using the entire dataset, suffering from insufficient ability to consider the differences among source patterns and lowering the predictability. To address this limitation, we propose the Multiple Time Series Pattern Representation Network(MTS-PRNet), a unified framework consisting of two modules to forecast multiple time series from diverse sources. The first is the intra-series correlation learning module, which explicitly learns the temporal dependencies of time series. The second is the inter-series discriminative representation learning module that learns shapelets as discriminative representations to capture shared features among series. By integrating the covariates map generated by the second module, both intra and inter-series characteristics are captured to provide transferable guidance for increasing predictability. Experiments conducted on 9 datasets verify that our model achieves state-of-the-art performance. In particular, we carry out an ablation study to validate the effectiveness of discriminative representations.