Paper ID: 2410.23749

LSEAttention is All You Need for Time Series Forecasting

Dizhen Liang

Transformer-based architectures have achieved remarkable success in natural language processing and computer vision. However, their performance in multivariate long-term forecasting often lags behind simpler linear baselines. Previous studies have identified the traditional attention mechanism as a significant factor contributing to this limitation. To unlock the full potential of transformers for multivariate time series forecasting, I introduce \textbf{LSEAttention}, an approach designed to address entropy collapse and training instability commonly observed in transformer models. I validate the effectiveness of LSEAttention across various real-world multivariate time series datasets, demonstrating that it not only outperforms existing time series transformer models but also exceeds the performance of some state-of-the-art models on specific datasets.

Submitted: Oct 31, 2024