Machine Learning Based Prediction and Optimization of Exergy Efficiency of Petroleum Refinery Under Uncertainty

Abstract

Energy efficiency studies are significant in the petroleum refinery because of their environmental impact and cost. In this context, crude distillation along with downstream units play an important role in refinery operations. This study established integrated frameworks that combine artificial neural networks (ANN) with particle swarm optimization (PSO) and genetic algorithm (GA). The objective was to achieve improved exergy efficiency in petroleum refinery operations under uncertainty. At first, a steady state Aspen HYSYS model was taken to execute the exergy analysis in order to measure the exergy efficiency, exergy destruction or irreversibility, and exergetic improvement potential of the overall plant model. The plant model was subsequently converted into a dynamic mode by initiating a ±10% uncertainty in the process parameters, such as pressure, temperature, and mass flow rates of 12 input streams. This resulted in the creation of a dataset consisting of 500 samples. Those datasets were utilized to create an Artificial Neural Network (ANN) model for the purpose of predicting the exergy efficiency. The ANN model was employed as a surrogate in Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) environments to get superior exergy efficiency in the presence of uncertainty. The optimum process condition obtained using GA and PSO approach were fed into the Aspen HYSYS model for validation. The steady state exergy efficiency of overall petroleum refinery model was 72.38%, while the irreversibility or exergy destruction and improvement potential of the overall plant wide model was 7311.97 kW and 2037.85 kW respectively. The ANN was trained using the scaled conjugate backpropagation (trainscg) training algorithm, with the network's activity being regulated by the tansig activation function. The RMSE was used to quantify the performance of the model architecture, having RMSE of 1.1349 for exergy efficiency. The R value for training is 0.99925, for validation is 0.93288, and for testing 0.91209. The performance of the particle swarm optimization (PSO) and genetic algorithm (GA) techniques were comparable, and they greatly improved the exergy efficiency of the overall plant compared to the steady state Aspen HYSYS model of the process.