基于数据增强技术与CNN-BiLSTM-Attention的油田注水流量预测及效果

准确识别地层注水情况是油田开发的重要前提，对制定合理的注水发展规划也具有重要的指导意义。为准确预测注水，提出一种结合卷积神经网络、双向长短期记忆网络与注意力机制的油田注水流量预测方法，该方法首先将卷积神经网络(convolutional neural network，CNN)和双向长短期记忆网络(bidirection long short-term memory，BiLSTM)进行联合，用于捕获注水流量的复杂非线性时空关系，然后采用注意力机制来关注输入的重要特征。并针对油田历史数据匮乏问题，提出使用数据增强技术来增加一维时间序列的数据量。采用国内某油田注水井真实历史注水数据进行实验，研究结果表明，本研究中提出的CNN-BiLSTM-Attention预测模型的平均绝对误差(Mmean absolute error，MAE)、均方根误差(root mean square errorRoot Mean Square Error, RMSE)、平均绝对百分比误差(mean absolute percentage errorMean Absolute Percentage Error，MAPE)和决定系数(Ccoefficient of Ddetermination，R2)MAE、RMSE、MAPE 和R2 分别为0.027、0.043、9.936和0.968，通过多种模型对比，表明该方法具有较高的预测精度，可以更准确地预测注水流量。此外，研究还证实，采用数据增强技术可以有效提高模型的预测精度。研究成果可为油田精细化注水提供调整方案与高质量数据，从而为油田智能化开发提供理论依据。

Oilfield Water Injection Rate Prediction and Effect Based on CNN-BiLSTM-Attention and Data Enhancement Technologies

Accurate identification of formation water injection is an important prerequisite for oilfield development and an important guide for making reasonable water injection development plans. To accurately predict water injection, an oilfield injection rate prediction method combining convolutional neural network, bi-directional long and short-term memory network and attention mechanism was proposed. The method first unites convolutional neural network(CNN) and bidirection long short-term memory(BiLSTM) for capturing complex nonlinear spatio-temporal relationships of water injection rate, then employs Aattention mechanism to focus on important features of the input. Data enhancement technologies were proposed to increase the amount of one-dimensional time series data in order to address the problem of insufficient historical data in oilfield. Experiments were conducted using real historical injection data from an oilfield injection well in China, and the results of the study showed that the Mmean absolute error(MAE)MAE, Rroot Mmean Ssquare Eerror(RMSE)RMSE, Mmean Aabsolute Ppercentage Eerror(MAPE)MAPE and Ccoefficient of Ddetermination(R2)R2 of the CNN-BiLSTM-Attention prediction model proposed in this study were 0.027, 0.043, 9.936 and 0.968, this method demonstrates higher prediction accuracy compared with various models and enables more precise estimation of the water injection rate. The study further demonstrated that the application of data enhancement techniques enhanced the predictive accuracy of the model. The research results can provide adjustment schemes and high-quality data for fine water injection in oilfield, thus providing a theoretical basis for the intelligent development of oilfield.