基于深度学习的油泡流聚并机制及含水率预测研究

Oil bubble flow widely exists in petroleum industry such as crude oil production and transportation. How to extract efficient features from large measurement bubble flow signals is always a problem to be solved. Deep learning theory provides a new way for solving this problem by extracting features hierarchically. In this study, we first collect large bubble flow fluctuating signals by using constructed multi-sensor measurement system. Then we extract efficient oil bubble features based on deep belief network and modeling the oil bubble coalescence phenomenon. In addition, based on the bubble coalescence model we investigate the oil bubble coalescence mechanism and reveal the flow pattern transition and evolution phenomenon. Also, we fuse the measured multi-sensor data with deep learning theory to set up a model for predicting the water phase fraction of low flow rate oil bubble flow. This model corrects the nonlinear relation between the measured water hold up and real water-cut. We try to use this model to solve the problem that water-cut is hard to predict under the condition of low flow rate bubble flow with high slippage velocity. This study develops the oil well producing profile testing theory which is significant to the redevelopment of oil fields.

油泡流是广泛存在于石油开采与储运过程中的流动现象。如何从大量实测油泡流信号中提取有效特征信息，并对油泡状态迁移及参数进行预测一直是亟待解决的问题。深度学习理论逐层抽象提取特征的方式为上述问题的解决提供了新的思路。本研究拟通过构建多传感器测量系统获取大量实测油泡流波动信号，基于深度置信网络提取油泡流有效特征信息。建立分散油泡聚并过程预测模型，研究低流速下油泡聚并模式及产生机制，认识油泡状态迁移及演化过程。同时，融合多传感器测量数据，建立基于深度学习的低流速油泡流含水率预测模型，修正油水相间滑脱导致的测量持率与真实含率间的非线性关系，解决低流速、高滑脱情况下泡状流含率预测困难问题。本项目的实施可丰富油井产液剖面测试理论，对油田二次开发具有重要意义。

目前，我国多数油田已进入高含水开发阶段。此时，产液地层的渗透特性发生了较大的变化，原油的产出形式由原来相分布均匀、流动平稳的状态转变为流速较低、相间作用复杂、产出及其不规律的油泡流。认识油井产出油泡流的流动规律，建立其参数测量模型一直是亟待解决的问题。针对上述问题，本项目从大量实测泡状流数据入手，将深度学习理论引入两相流测量领域，完成了以下研究内容：1）构建了基于深度卷积神经网络的流型识别模型，实现了油泡流的高精度辨识。2) 基于深度自编码网络建立了油泡流聚并检测模型，对垂直管道内的油泡流聚并现象进行了检测。3）结合油泡流的聚并检测结果及实验流动工况，研究了泡状流的聚并机制及流型演化机理。明确了流速、含水率以及流动形态与油泡流聚并机制的内在联系。4）提出了“两相流序数模式及序数复杂网络分析方法”，定量刻画了泡状流生成过程中的系统确定性变化规律，研究了由段塞流向泡状流转换的机制，采用多层序数复杂网络定量分析了泡状流的空间耦合特性。5）提出了提出了基于多元信息融合及深度LSTM网络的两相流参数预测方法，对低流速、高含水、高滑脱工况下的油泡流含率进行了预测，模型对垂直油水两相流含水率的预测误差达到1%以内。本项目的研究成果丰富发展了油井参数测试理论，项目实验数据，关键结论为我国陆上油田的生产优化管理及增产、增收策略提供了重要的指导资料。