基于长短期记忆神经网络模型的分层注水优化方法

Optimization of stratified water injection based on long-short term memory neural network model

分层注水是改善层间注采矛盾、提高水驱开发效果的一种重要手段。基于油藏数值模拟的分层注水优化存在地质模型不确定性强、所需数据多、计算耗时长等缺点,数据驱动的优化方法可有效克服上述缺点。以井组中所有注水井的分层注水层段为考察对象,采用平均不纯度减少(MDI)方法筛选影响每口生产井产液量和含水率的主要注水层段,以此为基础利用注水井分层注水量以及生产井产液量和含水率时序数据建立长短期记忆神经网络(LSTM)深度学习预测模型,结合粒子群优化算法(PSO)实现分层注水量优化。实例应用表明:基于注水井分层注水量的LSTM模型可以准确预测产液量和含水率,平均误差分别为0.5%和1.7%;在总注水量基本保持不变的情况下,优化后井组产油量增加12.2%、平均含水率下降4.2个百分点,实现较好的增油控水目的,为深度学习在分层注水优化方面的应用研究提供了一种新的方法。

Stratified water injection is an important means to improve the injection-production contradiction between layers and enhance the development effect of water flooding.The optimization of stratified water injection based on reservoir numerical simulation has some disadvantages,such as strong uncertainty of geological model,large amount of data required,and long calculation time.The data-driven optimization method can effectively overcome the above disadvantages.Taking the water injection intervals of all water injection wells into account,the main water injection intervals affecting daily production rate and water cut of each producing well were screened by means of Mean Decrease Impurity(MDI).On this basis,a Long-Short Term Memory(LSTM)deep-learning model was proposed with the time-series data of stratified water injection rate of water injection wells and liquid production rate and water cut of producing wells.At last,Particle Swarm Optimization(PSO)was used to optimize the stratified water injection rate.The field application indicates that the LSTM model based on the stratified water injection rate of water injection wells can accurately predict liquid production rate and water cut with average errors of 0.4%and 1.8%respectively.Under the condition of remaining the total water injection volume basically unchanged,the oil production of the well group increases by 12.2%and the average water cut decreases by 4.3%after optimization,realizing the purpose of increasing oil production and decreasing water cut.It provides a new method for the application of deep learning in the optimization of stratified water injection.