The selection of reservoir parameter evaluation network modeling based on serial or parallel multi task learning is a recent emerging issue.This article is based on experimental data of reservoir parameters in an old oilfield area in western China.Compare the evaluation results of porosity,saturation,and permeability of multi-task learning networks that are composed of 20 types different basic neural modules.This paper proposed selection strategies:the coefficient of determination as the evaluation index,the condition for selecting a serial rather than parallel structured multi-task learning network model is that the model parameter quantity is less than about 1000;When the number of model parameters is greater than 1000,serial multi-task networks are not as good as parallel multi-task networks.When the mean absolute error is the evaluation indicator,the prerequisite for selecting a serial multi-task network is a reference value with a model parameter quantity less than 10000.When the number of model parameters is greater than 10000,the results of serial and parallel multi-task networks have some similarity.If the mean absolute error and model parameter quantity are within the allowable range,both architecture networks are feasible.This paper aims to provide support for the design and application of different types of multi-task learning network architecture models.
NETL
NSTL
万方数据
