Predicating Water-flooding Performance into Stratified Reservoirs Using a data driven proxy model
分析注水动态时使用的数学模型通常基于假设，不考虑一种或多种力的影响，以应对油藏非均质性问题。本文采用敏感性分析研究了影响油藏分层注水动态的参数，这些因素旨在评估水驱过程中粘滞力、重力和毛细管力的影响。此外，还研究了由粘度和重力分离引起的窜流现象。对影响参数进行随机抽样，创建具有特定输入和目标输出的知识域，它是储层模拟器的最终采收率。该域用作代理模型（人工神经网络）的输入，用于在训练过程中调整神经元之间连接的大小，以生成一个模型，该模型可以在有限范围内预测此类油藏的注水效果，误差百分比很小。
该模型在提供12个关键参数（流度比、流体密度、倾角、渗透率顺序、非均质性程度、注入速度、储层厚度、三维孔隙度和渗透率以及储层深度）的情况下，可以预测非均质储层水驱过程的动态。当系统参数在训练过程中使用的数据范围内时，平均绝对误差百分比约为4.6%，误差标准差约为8.7%，模拟结果与人工神经网络的相关系数约为99.1%。
Abstract
Fundamentally, all mathematical models employed in analysis of water-flooding performance implied assumptions to exclude one or more forces to cope with the reservoir heterogeneity. In the beginning of the survey, a series of sensitivity investigations were undertaken to examine the parameters that affect the water-flooding performance in stratified reservoirs. The factors were designed to measure the impact of each force that contributed in water-flooding process. The forces are: viscous force, the force of gravity and capillary forces. Additionally, the cross flow phenomena which result from the viscosity and gravity segregation are investigated. The parameters that affected performance to a high degree were sampled randomly to create a knowledge domain with specific inputs and target outputs.
In this case, it was the final oil recovery factor by reservoir simulator tool. This domain is used as input (supplied solved problems) to the proxy model (artificial neural network) for adjusting the magnitude of the connections between the neurons during training process to generate a model that can predict the performance of the water-flooding in such reservoirs within a limited range with very minor percentage of error. This model can anticipate the performance of the water-flooding process in heterogeneous reservoir when supplied with 12 key parameters (mobility ratio, density of fluids, dipping angle, permeability ordering, heterogeneity degree, injection rate, reservoir thickness, porosity, and permeability in 3D and reservoir depth). The average absolute percentage of error is about 4.6% particularly and error standard deviation about 8.7% with correlation coefficient between result collected from simulation and ANN is about 99.1%, when the system parameters are within the range of data that was used during the training.
Key words: Secondary recovery techniques, water flooding, Neural Network, Stratified reservoirs
