Prediction of CO2storage performance in reservoirs based on optimized neural networks

Accurate prediction of CO2geological storage performance is of great guiding significance for oilfield development and management. The use of artificial neural networks is an effective method; however, insufficient data and certain errors caused by difficulties encountered during the measurement process make the traditional neural network fall into local optimization or overfitting. Therefore, in this study an optimized back propagation (BP) neural network based on the lion swarm optimization (LSO) algorithm was proposed. Firstly, the LSO algorithm was improved by combining a tent chaotic map and differential evolution to enhance its optimization capability. Secondly, the improved algorithm was used to obtain the optimal initial weights and thresholds of BP neural networks. Finally, by using the normalized sample data obtained by numerical simulation, five dimensionless variables were introduced to predict the CO2storage performance in reservoirs. Experimental results demonstrated that the proposed model yields a faster convergence speed and higher prediction accuracy compared with four existing neural networks. The root mean square errors on the training and test sets were 0.0234 and 0.0254, respectively, and the absolute error of more than 95% of the data was within 5%, which shows that the proposed method is feasible and effective in predicting the CO2storage performance and can serve as a good guide in oilfield development projects.