Positional encoding (PE) plays a critical role in Transformer-based time series forecasting. The study explores two newly proposed PEs: Temporal Position Encoding (T-PE) for temporal tokens and Variable Positional Encoding (V-PE) for variable tokens. These are designed to improve model robustness and forecasting accuracy.

Innovative Aspects of the Study:

• Novel Positional Encodings: T-PE and V-PE introduce geometric and semantic positional strategies, enhancing token relation comprehension.
• Dual-branch Framework: The T2B-PE framework efficiently calculates correlations using these new PEs, showcasing improved model performance in data-rich environments.

Importance of the Study: Aligning model capabilities with real-world data complexities is crucial for better forecasting outcomes. The enhancements in PE as presented in this research could lead to significant breakthroughs in AI’s ability to predict future events based on past data. It provides a foundational step toward refining the prediction models used in various real-time monitoring and decision-making tasks.