基于CNN-BiGRU混合神经网络的电潜螺杆泵产液量预测方法

为了解决电潜螺杆泵产液量虚拟计量的预测精度和稳定性，提出了一种基于双重注意力机制的卷积神经网络(CNN)和门控循环单元(GRU)混合模型(CNN-BiGRU)。油井产量预测易受油压、套压及泵工况等因素的影响，先利用皮尔逊相关系数和主成分分析法对数据进行降维，并确定主要影响因素；再利用CNN网络的局部连接和全局共享来提取历史油井产液量的空间特征；然后将特征送入GRU网络，提取数据的时间特征；最后利用双重注意力机制为不同的特征赋值对应的权值，进一步提升模型的预测精准度。将该方法应用到50 000条产液量数据样本上，模型预测精准度RMSE和MAPE分别取得5.24%、3.17%。与GRU、CNN-GRU模型相比，所提方法应用效果显著，能有效提升预测精度，具有一定的工程应用价值。

A CNN-BiGRU liquid production prediction method for electric submersible progressive cavity pumps

To improve the prediction accuracy and consistency of liquid production virtual metering of electric submersible progressive cavity pumps (ESPCPs), a hybrid model (CNN-BiGRU) integrating convolution neural network (CNN) and bidirectional gate recurrent unit (GRU) was proposed, which also incorporated the dual attention mechanism. This production forecast of oil wells is subjected to tubing pressure, casing pressure and pump conditions. First, the dimensionality reduction of data was performed using Pearson’s correlation coefficient and principal component analysis, and the main influential factors were identified. Subsequently, the local connecting and global sharing of the CNN network was used to extract spatial features of liquid production data of oil wells. The extracted features were then fed to the GRU network to extract the temporal features of the data. Finally, the weights were assigned corresponding features via the dual attention mechanism to improve the prediction accuracy of the model. The presented method has been applied to 50000 data samples of liquid production. The root-mean-square error (RMSE) and mean absolute percentage error (MAPE) of the model were 5.24% and 3.17%, respectively. Compared with the GRU and CNN-GRU models, the presented method delivers excellent prediction performance and greatly improves the prediction accuracy, which marks its high value for engineering applications.