A Multi-output Integration Residual Network for Predicting Time Series Data with Diverse Scales

Deep learning methods can fit the observation history over different time series with multiple levels of representations from huge dataset. However, it is challenging to directly train deep neural networks on a raw dataset with a large number of time series, as the different time-series have diverse scales. We initiate the study of an effective deep residual framework named MIR-TS for time series prediction with multi-output integration on time series data with diverse scales. Specifically, we leverage the residual module that constrains the original input average close to 0 to transform the original input, so that the distribution of features changes from sparse to dense. Compared with the traditional residual network, this approach improves the generalization of model via residual reuse, capturing more detailed features of time series to improve prediction. The results on the M3 and TOURISM benchmarks show that MIR-TS achieves a consistent better or highly comparable performance across different time series frequencies.