Leveraging Large Language Models for Discrepancy Value Prediction in Custody Transfer Systems: A Comparative Analysis of Probabilistic and Point Forecasting Approaches

Discrepancies in custody transfer systems in the oil and gas industry pose significant financial, regulatory, and operational risks. Accurate prediction of these discrepancies is critical to optimizing operations and minimizing potential losses. This study evaluates the effectiveness of Large Language Models (LLMs), specifically the Chronos-FineTuning Amazon Chronos T5 Small model, alongside statistical, machine learning, and deep learning models, in both probabilistic and point forecasting tasks. The evaluation covers metrics such as Weighted Quantile Loss (WQL), Scaled Quantile Loss (SQL), Mean Absolute Error (MAE), Symmetric Mean Absolute Percentage Error (SMAPE), and Root Mean Square Error (RMSE). The results highlight the superior performance of the Chronos model in both forecasting paradigms, demonstrating its ability to capture uncertainty and deliver precise predictions. This research offers valuable insights into selecting forecasting methodologies to improve custody transfer operations, underscoring the transformative potential of LLMs in industrial applications.