A dynamic patching LLM framework for adaptive yield trend prediction

Accurate yield trend prediction in petroleum refining is important for monitoring process evolution and supporting timely operational assessment. However, the dynamic and nonlinear nature of refining processes, together with the challenges of industrial data, complicates this task. Traditional rule-based systems lack flexibility, while deep learning methods often struggle with data noise and interpretability. We propose a dynamic patching LLM framework that adaptively constructs multi-scale temporal features, integrates process knowledge through semantic-guided prompts, and refines trend predictions via a self-evolving reasoning mechanism. The task predicts whether the target variable increases at each future step rather than estimating its continuous magnitude. Experiments show that our approach outperforms traditional deep learning and recent LLM-based baselines, demonstrating its effectiveness for multi-step trend classification across the evaluated refinery and public process datasets.