基于电阻抗谱特性与集成神经网络的天然气水合物饱和度软测量模型

Soft Sensing Model of Gas Hydrate Saturation Based on Electrical Impedance Characteristics and Ensemble Neural Network

摘要:

含天然气水合物饱和度的计算是储层优选和资源量评估的关键参数,针对目前数据解释模型计算精度低以及模型输入参数少等问题,提出了一种基于电阻抗特性参数和集成神经网络的软测量模型建立方法。在对电阻抗谱数据进行预处理、特征参数提取以及选择的基础上形成了样本集,针对四对传感器分别设计了BP神经网络,采用平均法作为集成策略将四个BP网络作为子网络进行集成得到集成网络模型。模型测试结果表明：通过集成网络模型计算得到的含水合物饱和度值平均相对误差3.33%、平均绝对误差0.0014、均方根误差为6.56%,三项误差指标均低于各个子网络的计算误差。在宽频范围内对含水合物沉积物进行电阻抗谱测试能够获得沉积物的频率响应特性以及特性描述参数,可为神经网络模型提供大量的输入参数；利用集成神经网络能够综合应用位于不同测量方位的多个传感器的测量数据,通过采用适合的集成策略能够克服水合物空间分布不均匀对饱和度计算准确度的不利影响。

Abstract:

The calculation of gas hydrate saturation is the key parameter of reservoir optimization and resource evaluation. In view of the low accuracy of data interpretation model and the lack of model input parameters, a soft sensing model establishment method based on electrical impedance characteristic parameters and ensemble neural network was proposed. On the basis of preprocessing, feature parameter extraction and selection of impedance spectrum data, a sample set was formed. BP neural network was designed for four pairs of sensors respectively. The average method was adopted to integrate the four BP networks and then an ensemble network model was obtained. It has been demonstrated that the average relative error of hydrate saturation calculated by the ensemble network model is 3.33%, the average absolute error is 0.0014, and the root mean square error is 6.56%. The three errors of the ensemble network are lower than those of each sub network. The frequency response characteristics and characteristic parameters of the hydrate-bearing sediment can be obtained from the electrical impedance spectrum test in a wide frequency range, which can provide a large number of input parameters for the neural network model. The ensemble neural network can be used to comprehensively apply the measurement data of multiple sensors located in different measurement directions. The adverse influence of non-uniform distribution of hydrate in space on the accuracy of saturation calculation can be overcome by adopting appropriate integration strategies.