A Related Work

A.1 Time Series Forecasting We first briefly review the related literature of time series forecasting (TSF) methods as below. Complex temporal patterns can be manifested over short-and long-term as the time series evolves across time. To leverage the time evolution nature, existing statistical models, such as ARIMA [6] and Gaussian process regression [7] have been well established and applied to many downstream tasks [28, 29, 2]. Recurrent neural network (RNN) models are also introduced to model temporal dependencies for TSF in a sequence-to-sequence paradigm [24, 9, 61, 40, 46, 50, 53]. Besides, temporal attention [49, 59, 56] and causal convolution [3, 5, 54] are further explored to model the intrinsic temporal dependencies.