INTELLIGENT CLUSTERING AND CLASSIFICATION OF THE GAS-CRUDE OIL SEPARATION PROCESS USING K-MEANS AND RANDOM FOREST: A DATA SCIENCE BASED APPROACH

Mejía R.D.V.; Rodríguez N.L.C.; Roca O.J.S.; Morales C.R. · Applied Engineering Letters (2025) · DOI: 10.46793/aeletters.2025.10.4.4

Resumen

Applying data science methodology, the study examined the separation of gas and crude oil in the state of the QE2 compressor in Monagas, Venezuela. It began with a descriptive statistical study that found and eliminated 2.22% of anomalous data, revealing a trimodal behavior for crude oil and a bimodal for gas. With skewness and a coefficient of determination (R2) where 0.7645 for the gas-crude ratio, both variables had a coefficient of variation greater than 20%. The K-means algorithm was used, which found four well-formed clusters. However, the Kruskal-Wallis method could not find statistically significant differences between them, suggesting that the variability is due to different operating rules, crude types or process errors, rather than clearly differentiated groups. Finally, a Random Forest algorithm was developed with one hundred trees. The most significant achieved an accuracy of 0.9929. Despite an initial Gini value of 0.725 (moderate impurity), it was segmented into two branches. The branch with a raw value ≤1.15 Thousands of Barrels of Crude Oil per Day (MBNPD) showed superior performance, with a Gini value of 0.01, indicating near-perfect purity. This shows that this branch classifies with high accuracy.