融合注意力机制的煤层气产量动态预测

为了提升煤层气产量的预测精度，提出融合注意力(Attention)机制并结合卷积神经网络（convolutional neural networks，CNN）和长短期记忆神经网络（long short term memory，LSTM)的煤层气产量动态预测模型。利用随机森林变量筛选方法，确定井底流压、动液面高度、套压、冲次为排采过程中影响煤层气产量的主控因素；利用CNN信息提取优势，提取煤层气排采数据的特征向量，并将特征向量作为LSTM网络的输入；再将LSTM隐含层融合注意力机制提取重要信息权重，有效解决信息长期依赖性和信息丢失。实验结果表明：融合注意力机制的CNN-LSTM煤层气产量动态预测模型各方面均表现较优。具体表现为：1. 模型预测性能较好，利用不同模型对比预测，改进后的煤层气产量预测模型精度最高，比标准的LSTM预测精度提升了3%~4%；2. 泛化性能较优，预测同一区块不同生产天数的6口煤层气井产量时，预测60天日产气量的平均相对误差均小于5%，预测200天日产气量的平均相对误差均小于8%。

Dynamic Prediction of Coalbed Methane Production by CNN-LSTM Integrating Attention Mechanism

In order to improve the prediction accuracy of coalbed methane production, a dynamic prediction model of coalbed methane production integrating attention mechanism, convolutional neural networks(CNN) and long short-term memory(LSTM) is proposed.The random forest variable screening method was used to determine bottom hole flow pressure,dynamic liquid level, casing pressure, and stroke is the main controlling factor affecting the production of coalbed methane in the process of pump and production; the CNN information extraction advantage could extract the characteristic vector of production data. And the feature vector was used as the input of the LSTM network; then the LSTM hidden layer integrating attention mechanism to extract important information rights, which could effectively solve the problem of long-term dependence and information loss. The experimental results show that all aspects of dynamic prediction of coalbed methane production by CNN-LSTM integrating attention mechanism model are better. Specific performance: 1. Model prediction performance is better, using different models to compare predictions, improved coalbed metallic production prediction model accuracy is the highest, 3% ~ 4% is increased than standard LSTM prediction accuracy; 2. The generalization performance of the model is better. When predicting the output of 6 coal methane wells with different production days in the same block, the average relative error of predicting 60 day daily gas production is less than 5%, and the average relative error of predicting 200 day daily gas production is less than 8%.