Investigating a User-Friendly Machine Learning Approach to Predict the Gas Hydrate Formation Temperature

ABSTRACT The petroleum industry faces a serious problem of gas hydrate formation in pipelines and process equipment, particularly in low-temperature marine environments. It is necessary to understand the chemical thermodynamics of gas hydrate formation so that it can be conveniently avoided. Although research has been conducted to develop correlations or to use Artificial Intelligence (AI) to determine hydrate temperature, to the authors’ knowledge, none have compared Machine Learning (ML) models’ performance in hydrate formation temperature prediction, especially Web User Interface (WUI) development based on the best-performing model. To achieve precision, this work compares the performance of four ML models, such as Artificial Neural Networks (ANNs), Decision Tree Regressor (DTR), Support Vector Regressor (SVR) and Random Forest Regressor (RFR) in terms of hydrate formation temperature prediction using operating pressure and specific gravity as features. Python was employed for this work, as it supports open-source libraries such as Keras with TensorFlow and scikit-learn, among others. Results showed that the coefficient of determination (R2), Root Mean Square Error (RMSE), and a20-index on the overall dataset are (0.9994, 0.3120, 1), (0.9992, 0.3716, 1), (0.9991, 0.3880, 1) and (0.9910, 1.2545, 0.9972) for DTR, ANNs, RFR, and SVR, respectively. Although all models delivered strong results, they differ sharply in the time required to train. ANNs required 974.9498 s, RFR needed 15.9753 s, and SVR took 294.4242 s. In contrast, DTR completed the task in only 0.2727 s. Based on performance and computational efficiency, the models rank as follows: DTR>RFR>SVR>ANNs. Eventually, Web User Interface (WUI) was developed based on the bestperforming model (DTR). The optimal activation function for the ANN is tanh, while the Support Vector Regressor model performs best with the Radial Basis Function (RBF) kernel. We are optimistic that this research will open novel avenues in natural gas engineering. It is recommended that the models’ lower and upper bounds be broadened by training the model on additional experimental data across different operating conditions.