Orthrus: A Dual-Branch Model for Time Series Forecasting with Multiple Exogenous Series.

Extending the forecasting horizon is a crucial demand for real applications in time series forecasting with multiple exogenous series (TFME). Previous studies adopt Transformer to effectively capture long-term dependency coupling between output and input in a sequence. However, the potential entanglement in multi-dimensional feature space still precludes the application in TFME tasks. In this paper, we propose a dual-branch network named Orthrus to solve this issue by differentiating the processing of target and exogenous sequences. Orthrus takes the long target sequence as input and uses multi-head self-attention mechanism to capture long-term cyclical patterns. Concurrently, it applies the Local Mutual Dependency Analysis module to extract the sub-sequences of the exogenous sequences with the maximum expected information and adopts the Multi-scale Convolutional Neural Network module to capture the dependencies among the sub-sequences and align with the target sequence. In this way, the feature entanglement issue is largely alleviated. Extensive experiments on four real-world datasets verify that Orthrus is superior to all baselines in prediction accuracy, inference efficiency, and memory usage, providing an effective solution to the TFME task.