Modelo predictivo basado en aprendizaje automático para la estimación de la compensación volumétrica por calidad (CVC) en sistemas de transporte de crudo en Colombia

Abstract

This work addresses the need to complement the Volumetric Compensation by Quality (CVC) process applied in oil pipelines, whose objective is to ensure the equitable distribution of transported crude oil volumes and values among the different system clients. Although this mechanism constitutes an essential tool for volumetric balance management, its traditional formulation presents several limitations: it relies on linear relationships, considers a limited number of variables (mainly API gravity and sulfur content), and operates reactively, without the ability to anticipate variations in operational conditions. To address this problem, a predictive model based on Machine Learning was developed to complement the traditional CVC method through the incorporation of artificial intelligence techniques. The DeepSets architecture was employed, which enables the processing of unordered and variable-sized data sets, thus better reflecting the real conditions of crude transport and mixing. In this context, the model was trained using historical volumetric balance data, where crude quality was not modeled directly from its physicochemical properties but rather through its relative economic value. This value synthesizes the combined effect of variables such as API gravity, viscosity, and sulfur content, reflecting their influence on market behavior and exchange agreements. Accordingly, the volumetric compensation was established based on value differences between input and output volumes, rather than on the individual physical properties of the crudes. The development process followed the CRISP-DM methodology, encompassing the stages of business understanding, data preparation, modeling, and predictive performance evaluation. The results demonstrate a high predictive capability of the model, achieving a determination coefficient of R2 = 0.96 and a mean absolute error (MAE) of 0.10, which indicates excellent correspondence between the predicted and actual volumetric compensation values. Furthermore, the model exhibited stability and robustness in the face of operational data variability, establishing itself as a reliable tool to estimate compensations and support logistical planning and volumetric balance validation. The proposed model does not aim to replace the current contractual CVC framework but rather to complement it through an automated and predictive approach that contributes to strengthening equity, transparency, and efficiency in the management of transported volumes. The integration of the DeepSets approach with the CRISPDM methodology represents a significant step toward the modernization of volumetric compensation processes, demonstrating the potential of Machine Learning to optimize operations within the energy sector.