PMTE-LLM:An LLM-based time series forecasting method using professional mechanism and training experience.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Time series data present complex patterns and noise, challenging deep learning models.
• Large language models (LLMs) are computationally intensive and memory-demanding.
• Existing model compression techniques often compromise accuracy.

Purpose of the Study:

• To develop a time series forecasting method that maintains accuracy while reducing computational costs.
• To address the limitations of current deep learning and LLM approaches in time series analysis.

Main Methods:

• Introduced a time series forecasting method based on LLMs, termed PMTE-LLM, integrating professional mechanisms and training experience.
• Employed multi-modal fusion to integrate time series data with knowledge texts and mechanism formulas into a unified feature space.
• Utilized a triangular mesh storage method for training, inspired by brain-like experience, and optimized parameters via reinforcement learning.

Main Results:

• PMTE-LLM reduced computational costs by 33%-54% and memory usage by 38%-65% compared to state-of-the-art models.
• Achieved accuracy improvements ranging from 1.8% to 47.3% across various tasks including classification, anomaly detection, and forecasting.
• In oil field operations, achieved over 97% production forecast accuracy with a 53% improvement in inference time efficiency.

Conclusions:

• The PMTE-LLM methodology effectively enhances time series forecasting accuracy and efficiency.
• The approach offers a superior alternative to existing methods, particularly for complex datasets and demanding applications.
• Demonstrated significant reductions in computational and memory overhead without sacrificing predictive performance.