Science Technology and Engineering
1671-1815
2022
22
18
7866
7872
article
基于PSO-GBDT的CO2-原油最小混相压力预测模型
Prediction Model of Minimum Miscible Pressure of CO2-Crude Oil Based on PSO-GBDT
随着“2060年碳中和目标”的提出，中国油田开发领域CO2气驱采油技术又一次得到了广泛关注。CO2气驱采油既可以实现对碳资源的地下封存，又能够对油田三次采油起到主要作用。但是其驱油效果会受到CO2和原油混相与否的制约，所以需要对CO2-原油体系最小混相压力（Minimum miscibility pressure，MMP）进行精准预测。而传统预测方法时间成本及误差过大，人工智能算法因其计算效率及准确率高便脱颖而出。本文使用随机森林算法对MMP主控因素进行分析，筛选出CO2, H2S, Cj, C2-C5, N2的摩尔分数及油藏温度、平均临界温度等特征变量，采用MLP、GA-RBF、RF、PSO-GBDT、AdaBoost SVR五种智能算法建立MMP预测模型。为此，本文使用了160行的数据库进行预测分析，采用五种不同评估指标及可视化图像对不同模型结果进行对比分析，并验证模型的准确性。最终测试效果证明，在数据有限的情况下，PSO-GBDT模型具有最佳的MMP预测效果，PSO-GBDT平均绝对百分比误差（Mean absolute percentage error，MAPE）为4.89%，均方根误差（ Root mean square error，RMSE）为0.83，测试集R2为0.96。此模型精度最高，灵活性、鲁棒性最强。
With the proposal of "carbon neutralization goal in 2060", CO2 flooding technology has attracted wide attention again in the field of oilfield development in China. CO2 flooding can not only realize underground storage of carbon resources, but also play a major role in tertiary oil recovery. However, its oil displacement effect is restricted by the miscibility of CO2 and crude oil, so it is necessary to accurately predict the minimum miscibility pressure (MMP) of CO2-crude oil system. Due to the large time cost and error of traditional prediction methods, artificial intelligence algorithm stands out because of its high computational efficiency and accuracy. In this paper, random forest algorithm was used to analyze the main control factors of MMP, and the mole fraction of CO2, H2S, Cj, C2-C5, N2, reservoir temperature, mean critical temperature and other characteristic variables were screened out. Five intelligent algorithms including MLP, GA-RBF, RF, PSO-GBDT and AdaBoost SVR were utilized to establish the MMP prediction model. The final test results show that the PSO-GBDT model has the best MMP prediction effect under the condition of limited data. The mean absolute percentage error (MAPE) of PSO-GBDT is 4.89%, the root mean square error (RMSE) is 0.83, and R2 of the test set is 0.96. This model has the highest accuracy, flexibility and robustness.
人工智能；最小混相压力；MMP预测；PSO-GBDT ；AdaBoost SVR
Artificial intelligence；Minimum miscible pressure；MMP prediction；PSO-GBDT ；AdaBoost SVR
沈斌,杨胜来
Shen Bin, Yang Shenglai
jsygc/article/abstract/2109674