Appendix

Each data point includes an "oil temperature" value and6power load features. This is because we could further extract essential information from coarser temporal resolutions with deeper levels in the SCINet whenT is large. From Table 9, we can observe a clear trade-off betweenL and K. Moreover, the performance variation under differentT also indicates the importance of the look-back window selection for forecasting tasks. LossFunction Toenhancetheperformance insingle-step (short-term timeseriesforecasting Sec.4.2)forecasting, we revise the loss function of the last SCINet in the stacked SCINet withK(K 1).