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

Abstract

Time series forecasting has seen extensive application including traffic and energy. Time series data often contain complex patterns and noise, posing challenges for deep learning models to effectively extract features and achieve accurate forecasting. Large language models (LLMs) suffer from high computational complexity and memory usage. Existing model compression strategies often lead to accuracy loss and performance degradation. This paper introduces a time series forecasting method based on LLMs driven by a mixture of professional mechanism and training experience (PMTE-LLM), aiming to maintain accuracy while lowering computational costs. By integrating time series data with knowledge texts and mechanism formulas through multi-modal fusion, PMTE-LLM maps different modalities into unified feature space. It helps LLM facilitate deeper understanding of underlying temporal dynamics. Inspired by human brain-like experience effects, triangular mesh storage method for training is introduced. This method adapts to the characteristics of input data to form specialized training experiences for different tasks. Parameters are continuously updated by reinforcement learning, dynamically optimizing weight matrices and computation paths. Experiments were conducted on various publicly datasets and oil production scenarios, covering tasks such as classification, anomaly detection, and forecasting. The results demonstrate that compared to state-of-the-art models from the past two years, PMTE-LLM approach reduces computational costs by 33 %-54 % and decreases memory usage by 38 %-65 %. Accuracy improvements ranged from 1.8 % to 47.3 % across various tasks. In the deployment within oil field operations, production forecast accuracy exceeded 97 %, with inference time efficiency improving by 53 %. These effectively showcase the efficacy and superiority of our proposed methodology.