A Novel Methodology for Predicting the Production of Horizontal CSS Wells in Offshore Heavy Oil Reservoirs Using Particle Swarm Optimized Neural Network

Cyclic steam stimulation (CSS) is one of the main offshore heavy oil recovery methods used. Predicting the production of horizontal CSS wells is significant for developing offshore heavy oil reservoirs. Currently, the existing reservoir numerical simulation and analytical models are the two major methods to predict the production of horizontal CSS wells. The reservoir numerical simulation method is tedious and time-consuming, while the analytical models need many assumptions, decreasing models’ accuracy. Therefore, in this study, a novel methodology combining the particle swarm optimization algorithm (PA) and long short-term memory (LM) model was developed to predict the production of horizontal CSS wells. First, a simulation model was established to calculate the cumulative oil production (COP) of horizontal CSS wells under different well, geological, and operational parameters, and then the correlations between the calculated COP and parameters were analyzed by Pearson correlation coefficient to select the input variables and to generate the initial data set. Then, a PA-LM model for the COP of horizontal CSS wells was developed by utilizing the PA to determine the optimal hyperparameters of the LM model. Finally, the accuracy of the PA-LM model was validated by the initial data set and actual production data. The results showed that, compared with the LM model, the mean absolute percentage error (MAPE) of the testing set for the PA-LM model decreased by 4.27%, and the percentage of the paired points in zone A increased by 2.8% in the Clarke error grids. In addition, the MAPEs of the training set for the PA-LM and LM models stabilized at 267 and 304 epochs, respectively. Therefore, the proposed PA-LM model had a higher accuracy, a stronger generalization ability, and a faster convergence rate. The MAPEs of the actual and predicted COP of the wells B1H and B5H by the optimized PA-LM model were 8.66% and 5.93%, respectively, satisfying the requirements in field applications.