Kinematic Viscosity of Petroleumat at 37.78 ℃: A Comprehensive Comparison of Machine Learning Techniques

The viscosity of crude oil is a significant component influencing oil recovery and flow behavior, yet accurately predicting viscosity remains a significant challenge in reservoir optimization, requiring the use of predictive models. This study evaluates eight Artificial Intelligence (AI) models, Feed-forward Neural Network (FFNN), Cascade Forward Neural Network (CFNN), Elman Neural Network (ENN), Multi-layer Perceptron Neural Network (MLPNN), Radial Basis Function Neural Network (RBFNN), k-Nearest Neighbor (kNN), Support Vector Regression (SVR), and Extreme Learning Machine (ELM), as well as three mathematical models, Poisson Regression Model (PRM), Quadratic Model (QM), and Multiple Linear Regression (MLR), for predicting the Kinematic Viscosity (KV) of crude oil at 37.78 ℃. A dataset of 274 crude oil samples was compiled from literature sources, including Molecular Weight (MW), Refractive Index (RI), Sulfur content (S),Specific Gravity (SG), and Initial Boiling Points (IBP)ranging from 70℃ to 565℃. To identify the most influential factors affecting KV, Principal Component Analysis (PCA), Pearson's Correlation Coefficient (PCC), and Decision Trees (DT)were applied, highlightingIBP-280 ℃, IBP-343 ℃, S, RI, MW, and SG as key variables. Subsequently, 63 MLPNN models were trained with various input combinations to evaluate their impact on prediction accuracy. Four models, M#1 [SG, MW], M#2 [SG, IBP-280 ℃, IBP-343 ℃], M#3 [SG, S, IBP-280 ℃, IBP-343 ℃], and M#4 [SG, IBP-280 ℃, IBP-343 ℃, RI, MW], demonstrated superior performance based on multiple statistical metrics. These four feature sets were then incorporated into the AI and mathematical models for further evaluation. Comparative analysis revealed that ELM, RBFNN, and PRM models exhibited the highest accuracy and stability among individual approaches. To further enhance prediction quality, a hybrid ensemble model (RBFNN-ELM-PRM) was developed, integrating the strengths of these three models. The ensemble, using the feature set [SG, IBP-280℃, IBP-343℃, RI, MW], outperformed all individual models, offering improved robustness and precision in KV prediction.