Leveraging Temporal Fusion Transformers and Domain-Specific LLMs for Real-World Industrial Sensor Forecasting and Decision Support

Abstract

In process industries, large volumes of sensor data are generated continuously, yet much remains underutilized for proactive decision-making. This thesis explores a novel architecture that combines deep learning and large language models (LLMs) to forecast, interpret and prevent process threshold violations in an industrial process facility. A Temporal Fusion Transformer (TFT) model was trained on 3 months of real-world, multivariate sensor data (1-minute resolution across 31 sensors) to predict 12-minute-ahead process parameter exceedances. Forecast outputs were passed to a costume-built domain-specific GPT-4.1 model, configured using prompt engineering, graph interpretation capabilities, and a retrieval-augmented generation (RAG) system incorporating expert literature and process knowledge. The GPT model synthesized probabilistic forecasts into well-structured team-based Five Whys root cause analyses, where virtual domain experts questioned each other to refine the diagnosis, a long term mitigation plan to remove the root causes found, and simulation-driven, per-unit prevention plan, generated by testing alternative process settings with the trained deep learning model and selecting the minimal production disturbance configuration that prevented the predicted violation, all while leveraging domain-specific knowledge to ensure operational feasibility and engineering trustworthiness by explicitly referencing authoritative sources from its RAG library, such as procedures and technical text books, to maintain compliance with stakeholder needs. Evaluation showed that GPU-trained deep learning model significantly outperformed CPU-trained equivalents in mean quantile loss metrics. Subject matter expert evaluation of the LLM’s responses indicates that the LLM’s insight quality improved as more domain knowledge was added leading to greater specificity, unit level differentiation in recommendations. This dual-model system demonstrates a scalable approach to combining forecasting and interpretability in one pipeline, offering preventative, actionable, domain-specific support for engineers, operators, and managers in complex industrial environments.