基于改进深度强化学习的注采调控模型研究

Research on Injection-Production Control Model Based on Improved Deep Reinforcement Learning

扫码查看

原文链接

维普
万方数据

中文摘要：提出一种基于改进深度强化学习的注采调控模型.首先,建立以最大化经济效益为目标函数的注采调控强化环境.其次,针对模型参数量大、网络内部协变量位移多等问题,提出一种改进双重深度Q网络的深度强化学习方法,应用批量归一化技术逐层归一化模型的输入数据,增强模型的泛化能力;再通过剪枝模块压缩模型体积,加速网络的训练过程,引入动态ε策略思想提高模型的鲁棒性和稳定性.最后,将所提模型同其他模型进行对比,实验结果表明,所提模型能获得更高、更稳定的平均累计奖励和更快的收敛速度和运行速度.

外文摘要：An injection production regulation model based on improved deep reinforcement learning is proposed. Firstly,es-tablish an injection and mining regulation and strengthening environment with maximizing economic benefits as the objective function. Secondly,aiming at the problems of large number of model parameters and many covariate displacements within the network,a deep reinforcement learning method to improve the double deep Q network is proposed,and the input data of the model is normalized layer by layer by applying batch normalization technology to enhance the generalization ability of the model,and then the model volume is compressed through the pruning module to accelerate the training process of the network,and the dynamic ε strategy idea is introduced to improve the robustness and stability of the model. Finally,the proposed model is compared with other models,and the experimental results show that the proposed model can obtain higher and more stable average cumulative reward,faster convergence speed and running speed.

外文关键词：

injection-production controldeep reinforcement learningpruningbatch normalizationdouble deep Q network

作者：

陈锐、张强、曾俊玮

展开 >

作者单位：

东北石油大学计算机与信息技术学院,大庆 163318

关键词：

注采调控深度强化学习剪枝批量归一化双重深度Q网络

基金：

国家自然科学基金黑龙江省自然科学基金黑龙江省博士后专项黑龙江省省属本科高校基本科研业务黑龙江省优秀青年教师基础研究支持计划

项目编号：

42002138LH2022F008LBH-Q200772022TSTD-03YQJH2023073

出版年：

2024

长春理工大学学报(自然科学版)

长春理工大学

长春理工大学学报(自然科学版)

CSTPCD

影响因子：0.432

ISSN：1672-9870

年,卷(期)：2024.47(3)

参考文献量12