基于K折双循环神经网络的缝洞型油藏单井注气效果预测方法

A method to predict single well gas injection effect for fractured-vuggy reservoirs based on Kfolding double recurrent neural network

塔河油田缝洞型碳酸盐岩油藏单井注气开发时各井效果差异较大,进行注气效果预测时注气井数达不到常规BP神经网络方法的要求。提出了K折双循环神经网络方法,该方法能够充分发挥样本数据的训练潜能,解决小样本条件下的神经网络训练难题;运用主成分分析法筛选出影响塔河油田单井注气效果的主控因素为底水能量、储量规模、注气量、储集体类型、采出程度等5项,以主控因素为输入参数,以注气增油量为输出参数构建了预测模型,对单井注气效果进行了定量预测。结果表明:本文方法预测注气后增油量值与实际值相比,误差最小为2.05%,最大为9.64%,符合工程上精度要求。本文方法为塔河油田缝洞型油藏注气选井提供了1种有效的定量分析工具,具有矿场应用价值。

During the single well gas injection development of fractured vuggy carbonate reservoirs in the Pahe oil field, distinctive development effects have been obtained in wells and the number of wells cannot meet the require ments of the regular BP neural network method for the gas injection effects prediction. Therefore, the K folding double recurrent neural network method was put forward, which can fully achieve the training potential of sample data and solve the problems about the training of limited samples; further, five main control factors for single well gas injection effects of the Pahe oilfield have been screened out by the main components analysis method including the bottom water energy, reserve scale, gas injection, reservoir type, and recovery, and a prediction model has been constructed with the main control factors as input parameters and the oil increment by gas injection as the out put parameter, used for the quantitative prediction of single well gas injection effects. Analysis results indicate that by comparing the oil increment value between the prediction with the method and the field application, the mini mum error is obtained of 2.05 % and the maximum error 9.64%, meeting the engineering precision demands. In conclusion, the method in this paper functions as an effective quantitative analysis tool for well selection aiming for gas injection in fracture＆vuggy reservoirs in the Tahe oilfield, being of site application value.