Prediction of Mercury Removal Efficiency of Commercial Adsorbent in Natural Gas Based on PSO-BP Neural Network

To reduce the loss of mercury adsorbents in natural gas and achieve accurate prediction of mercury removal efficiency, the mercury removal efficiency of commercial adsorbent was predicated using various models in the present study. The mercury removal experiment of commercial adsorbent was carried out in simulated nature gas and 304 sets of experimental data were obtained. Regression prediction models including Random Forest (RF), Support Vector Regression (SVR), Stochastic Gradient Descent (SGD), Ridge Regression (RR), and BP neural network (BPNN) were employed to predict the mercury removal efficiency of commercial adsorbent. The parameters of temperature, gas flow velocity, mercury adsorbent dosage, removal efficiency, and reaction duration were set as inputs, whereas the current removal efficiency was defined as output. The BPNN was the optimal model to predict the mercury removal performance of commercial adsorbent in nature gas, because of its high $R^{2}$ value (0.9219) and low RMSE (0.0445) and MAPE values (5.37%). To further improve the prediction accuracy, the particle swarm optimization algorithm (PSO) was introduced to optimize the hyperparameters of the BPNN. After optimization, the $R^{2}$ improved to 0.9532, while the values of RMSE and MAPE decreased to 0.0345 and $4.3 \%$, respectively. According to the above results, the PSO optimized BPNN model might prove to be a promising model to predict mercury removal efficiency of commercial adsorbent in natural gas.