Estimation of the methanol loss in the gas hydrate prevention unit using the artificial neural networks: Investigating the effect of training algorithm on the model accuracy

Artificial neural networks (ANNs) can understand the behavior of a given system from the historical measurements of its associated variables. Adjusting the weight and bias of the ANN model using an optimization algorithm is known as the training process. The ANN reliability is directly related to the success of the training process. Therefore, this study investigates the effect of optimization algorithms on the prediction accuracy of the multilayer perceptron neural networks (MLPNNs). The complex gas hydrate prevention unit is simulated using the MLPNN model trained by 20 different optimization algorithms. This study investigates the gradient-based, evolutionary, and Bayesian-based optimization algorithms. Combining statistical and ranking analyses confirms that the Levenberg–Marquardt (LM) is the most efficient optimization technique for training the MLPNN model. This training algorithm adjusts the weight and bis parameters of the MLPNN so that the highest accurate predictions have been achieved. On the other hand, the trained MLPNN by imperialist competitive algorithm shows the lowest accuracy for the considered task. The trained MLPNN by the LM algorithm predicts 239 laboratory-measured data sets about the methanol (MeOH) loss with the absolute average relative deviation of 6.4% and regression coefficient of 0.9643. Coupling the developed MLPNN and differential evolution optimization algorithm shows that temperature = 263 K and pressure = 6.92 MPa are the optimum condition for minimizing the MeOH loss in the gas hydrate prevention unit. Economic analysis confirms that the annual cost of methanol loss for the daily processing of 100 × 10⁶m³ of gas is ~17 million US dollars.