In this paper, we intent to handle this emerging issue.

The feed-forward hidden dimension of Transformers is set to 1024. AC-to-all Transformer decoders, have 2 layers. The same setup is used for both the WOMD dataset and the A V2 dataset. In the following, we report the configuration of ablation models. VRAM (RTX 3090 in our case) because they require more GPU memory at training time.

Time series analysis is widely used in extensive areas.

Model Category Key Characteristics iTransformer [11] Transformer-based Processes each variate independently prior to multivariate fusion and is regarded as the current state-of-the-art (SOTA) in time series forecasting. PatchTST [42] Transformer-based Divides the time series into patches and applies channel-independent shared embeddings and weights for feature extraction. Crossformer [35] Transformer-based Utilizes cross-attention mechanisms to effectively capture long-range dependencies across temporal sequences.FEDformer [43] Transformer-based Improves Transformer performance by leveraging frequency-domain sparsity, typically through Fourier transforms. Autoformer [33] Transformer-based Employs a decomposition-based architecture combined with an auto-correlation mechanism for effective time series modeling. RLinear [44] Linear-based A state-of-the-art linear model that incorporates reversible normalization and assumes channel independence.TiDE [45] Linear-based An encoder-decoder architecture built entirely using multi-layer perceptrons (MLPs). DLinear [46] Linear-based Among the earliest linear models for time series forecasting, utilizing a single-layer architecture combined with series decomposition. TimesNet [28] Temporal Conv-based Employs 2D convolutional kernels (TimesBlock) to model both intra-period and inter-period variations in time series data.