基于遗传算法的最小二乘支持向量机预测凝析气藏露点压力

露点压力的准确预测对保障凝析气藏的高效开发至关重要。近年来,数据挖掘、人工智能等大数据技术逐渐成为研究热点,其对复杂的非线性回归与分类问题有良好的解决策略。基于优化算法和机器学习,提出了一种将遗传算法(GA)与最小二乘支持向量机(LSSVM)相结合的露点压力预测模型(GA-LSSVM模型),并利用误差反向传播(BP)和径向基函数(RBF)人工神经网络建立了相应的露点压力模型,然后进行模型精度对比。在皮尔逊关联性分析基础上,上述模型均选取气藏温度、(N_2+CO_2、C_1、C_2～C_6、C_(7+))摩尔分数、C_(7+)相对分子质量、C_(7+)相对密度和气油比作为自变量。采用公开发表的34个露点压力数据进行参数优化,得到了GA-LSSVM、BP和RBF模型的最优参数,并对15组实测露点压力数据进行预测。结果表明:GA-LSSVM模型预测精度明显高于BP、RBF神经网络模型,具有良好的预测能力,GA-LSSVM模型的平均绝对相对误差(AARD)仅为3.02%,其中最大绝对相对误差(ARD)为16.64%,最小ARD为0.05%,BP和RBF神经网络模型的AARD分别为6.46%、10.54%。最后,根据Leverage方法,进行了所有数据的异常点检测。研究为凝析气藏露点压力预测提供了一种有效方法。

Dew point pressure prediction of condensate gas reservoir based on GA-LSSVM

The accurate prediction of dew point pressure is very important to ensure the efficient development of condensate gas reservoirs. In recent years, data mining, artificial intelligence and other data technologies have gradually become a research hotspot, which has a good strategy for solving complex non-linear regression and classification problems. Based on optimization algorithm and machine learning, a dew point pressure prediction model (GA-LSSVM) is proposed, which combines genetic algorithm (GA) with least squares support vector machine (LSSVM). Back propagation (BP) and Radial Basis Function (RBF) artificial neural networks are used to establish the corresponding dew point pressure model, and then the accuracy of the model is compared. Based on Pearson correlation analysis, gas reservoir temperature, (N2+CO2, C1, C2-C6, C7+) molar fraction, C7+ relative molecular mass, C7+ relative density and gas-oil ratio are selected as independent variables in the above models. The optimal parameters of GA-LSSVM, BP and RBF models were obtained by optimizing the parameters of 34 dew point pressure data published publicly, and 15 groups of measured dew point pressure data were predicted. The results show that the prediction accuracy of GA-LSSVM model is obviously higher than that of BP and RBF neural network models. The average absolute relative error (AARD) of GA-LSSVM model is only 3.02%. The maximum absolute relative error (ARD) is 16.64%. The minimum ARD is 0.05%. The AARD of BP and RBF neural network models is 6.46% and 10.54%, respectively. Finally, according to Leverage method, anomaly detection of all data is carried out. This study provides an effective method for predicting dew point pressure in condensate gas reservoirs.