基于迁移学习的地球物理测井储层参数预测方法研究

随着大数据和机器学习的成熟和推广应用，人工神经网络在地球物理测井预测储层参数中得到重视.本文引入迁移学习进行测井储层参数预测，以孔隙度预测神经网络模型和孔隙度含水饱和度联合预测神经网络模型为基础模型，分别以渗透率及含水饱和度预测作为目标任务进行迁移学习，以提升储层参数预测效果和效率.文中详细阐述了基于迁移学习的测井储层参数预测方法，并使用64口井的测井数据进行储层参数预测效果分析.结果表明，使用迁移学习后，渗透率模型预测效果最高可以提升58.3%；含水饱和度模型预测效果最高可以提升近40%，且最大可以节省60%的计算资源；以孔隙度预测模型为基础模型时更适合使用参数冻结的训练方式，以孔隙度含水饱和度联合预测模型为基础模型时更适合使用参数微调的训练方式.

A reservoir parameters prediction method for geophysical logs based on transfer learning

Geophysical logging is the main measurement to evaluate reservoir parameters, such as porosity, permeability and saturation. Generally, the reservoir parameters prediction with logging data is based on petrophysical model and response functions, which requires clear relationship and mechanism. However, in the actual underground detection, there are strong uncertainty of the response mechanism of electrical, acoustic, nuclear radiation and nuclear magnetic resonance methods, which makes difficulty of the application of petrophysical mechanism model. With the maturity and extension of big data and machine learning, more researchers have paied attention to the application of artificial neural network in the reservoir parameters prediction with logs. It could fiind the mapping between geophysical logging data and reservoir parameters by constructing a suitable network model and training with high-quality labeled data, without the domain knowledge of geology, geophysics and petrophysics. In this paper, transfer learning is introduced to the reservoir parameters prediction with logs. Taken porosity prediction network model and porosity water saturation joint prediction neural network model as base models, permeability and water saturation prediction as target models, we use transfer learning to improve prediction performance and training efficiency. We also suggest a method of constructing transfer learning neural network model for reservoir paramenters prediction, and analysis the performance with logs from 64 wells. The preliminary results show that, in the best cases, the prediction performance of permeability model using transfer learning can be increased by 58.3%; the prediction performance of water saturation model using transfer learning can be increased by nearly 40%, and the calculation resources can be saved by 60%; the transfer parameters freezing training mode is more suitable for the porosity prediction model based transfer learning model, and it is more suitable to use the transfer parameters fine-tuning training mode for the porosity and water saturation joint prediction model based transfer learning model. In the future, it could be considered to use instance-based transfer learning solving the problem of simple samples; use feature-based transfer learning to reducing the impact of poor-quality labels on the model.