Abstract
High-dimensional time series data forecasting has been a popular problem in recent years, with ubiquitous applications in both scientific and business fields. Modern datasets may incorporate thousands of correlated time series that evolve together, and correctly identifying the correlated patterns and modeling the inter-series relationship can significantly promote forecast accuracy. However, most statistical methods are inadequate for handling complicated time series due to violation of model assumptions, and most recent deep learning approaches in the literature are either univariate (not fully utilizing inter-series information) or computationally expensive. This paper present SM-TCN, a Sparse Multi-scale Temporal Convolutional Network, utilizing a forward–backward residual architecture with sparse TCN kernels of different lengths to extract multi-resolution characteristics, which sufficiently reduces computational complexity specifically for high-dimensional problems. Extensive experiments on real-world datasets have demonstrated that SM-TCN outperforms state-of-the-art approaches by 10% in MAE and MAPE, and has the additional advantage of high computation efficiency.