基于自动机器学习的采油井压裂效果预测方法

Prediction method for hydraulic fracturing effect of oil production well based on automatic machine learning technology

目前大庆油田采油井压裂效果预测时多是凭借经验或者多元线性回归等简单模型,存在着预测结果稳定性差且预测精度不高的问题。以大庆油田N23区块为例,借助数理统计方法对采油井压裂效果与各项影响因素开展了相关性分析,并采用随机森林算法研究了各影响因素对N23区块采油井压裂效果的影响程度;阐述了自动机器学习中元学习、贝叶斯优化和模型集成这三项关键技术的原理以及实现方法,并利用自动机器学习建立了基于数据驱动的采油井压裂效果预测模型;同时,将自动机器学习预测模型与随机森林、支持向量机和神经网络这3种常见机器学习算法的预测性能进行了对比,并利用该自动机器学习预测模型对N23区块的水力压裂进行设计与优化。结果表明,压裂前的生产参数对预测采油井压裂效果有着重要的影响;自动机器学习预测模型比其他算法的精度更高,模型在测试集上的决定系数为0.695,预测结果相对误差的平均值为18.96%,比目前水平降低了57.53%;44经过模型优化的压裂方案较原方案增加经济效益约3.2×10^(4)~27.4×10^(4)元/井次。

At present,the prediction of the hydraulic fracturing effect of oil production wells in Daqing Oilfield mostly re⁃lies on experience or simple models such as multiple linear regression,which leads to poor stability of prediction results and low prediction accuracy.With Block N23 of Daqing Oilfield as an example,the correlation between the fracturing effect of oil production wells and influencing factors is analyzed by the mathematical statistics.The influence of those factors on the hydraulic fracturing effect in Block N23 is studied by a random forest algorithm.Additionally,the principles and imple⁃mentation methods of meta learning,Bayesian optimization,and model ensemble in automatic machine learning are present⁃ed,and a prediction model of a data-driven hydraulic fracturing effect based on the automatic machine learning technology is constructed.Meanwhile,the model is compared with three common machine learning algorithms:random forest,support vector machine and neural network.The proposed model is employed to design and optimize the hydraulic fracturing of Block N23.The results show that the production parameters before fracturing exert an important influence on predicting the effect of oil production well after fracturing.The model constructed by the automatic machine learning algorithm has higher accuracy than other algorithms.The determination coefficient on the test set is 0.695,and the average relative prediction er⁃ror is 18.96%,which is 57.53%lower than the current level.Compared with the original one,the fracturing scheme optimized by the model can increase the economic benefit by about 3.2×10^(4)~27.4×10^(4) yuan per well.