CBiA-PSL抽油井异常工况预警模型

Early warning model for abnormal working conditions of CBiA-PSL pumping wells

油田生产过程中,油井受各种因素的影响容易发生泵漏、管漏等异常工况,会降低油井产出甚至导致躺井,对异常工况预警是油田智能化管理的重要任务。基于CNN-BiGRU联合网络,提出一种改进的网络结构CBiA-PSL模型(CNN BiGRU attention-positive sharing loss),用于油井异常工况预警。模型利用CNN学习工况样本灰度图像的深度特征,BiGRU有效避免信息损失并加强CNN池化层特征的联系,注意力机制对隐藏状态加权计算以完成有效特征筛选。针对工况数据集不平衡的问题,提出正共享损失函数PSL,将异常数据(正类)划分为子类,每个子类都共享整个正类的损失,且给样本少的正类更高的权重。实验结果表明,CBiA-PSL模型预测效果更佳,对于异常类和整体的预测都有较高的精度。

In the process of oilfield production,which is affected by various factors,oil wells are prone to abnormal working conditions,such as pump and pipe leakages,which will reduce the output of oil wells and even lead to lying in wells.Early warning of abnormal working conditions is an important task of intelligent oilfield management.Based on the convolutional neural network-bidirectional gated recurrent unit(CNN-BiGRU)joint network,an improved network structure CNN-BiGRU attention-positive sharing loss(CBiA-PSL)model is proposed for early warning of abnormal oil well conditions.The model uses the CNN to learn the depth features of the gray image of the sample,BiGRU to effect-ively prevent information loss and strengthen the connection between the features of the CNN pool layer,and attention mechanism to weigh the hidden state to complete effective feature screening.To address the imbalance of the working condition dataset,a PSL function,which divides the abnormal data(positive class)into subclasses,is proposed.Each subclass shares the loss of the entire positive class and provides a high weight to the positive class with only a few samples.The experimental results show that the CBiA-PSL model has a better prediction effect and a higher accuracy for anomaly and overall prediction than other models.