煤层气井产气量是衡量一口煤层气井产气能力和工程开发效果的重要指标，准确预测日产气量是保证煤层气高效生产的一个关键问题。以沁水盆地南部郑庄区块3号煤层为研究对象，选取煤层气井排采动态参数，如井底流压、液柱高度、套压、日产水量和冲次作为自变量，分析了日产气量与这些排采参数之间的相关性，建立了基于循环神经网络的煤层气井产气量预测模型与方法。研究结果表明，煤层气井日产气量与冲次呈正相关性，日产气量与井底流压、套压、液柱高度和日产水量呈负相关性。基于深度学习随机森林算法中的特征重要性分析，研究了排采动态参数与日产气量之间的非线性关系以及预测模型中对日产气量的贡献率，得到了排采参数对日产气量影响的重要性排序表现为：井底流压>液柱高度>套压>日产水量>冲次。在此基础上，基于循环神经网络改进的长短时记忆神经网络预测模型，将Z4-026井排采数据代入模型计算，预测了煤层气井未来60 d产气量情况，并将预测结果与传统的支持向量机回归模型、随机森林回归模型以及BP神经网络模型对比，发现改进的长短时记忆神经网络预测模型，拟合效果相对较好，实际日产气量与预测日产气量之间的误差小于5％。在郑庄区块5口煤层气井的产气量预测分析中，相对误差小于10%。因此该方法将为煤层气井产气量预测和制定合理的排采制度提供了有效途径。

The gas production of a coalbed methane well is an important indicator to measure the gas production capacity of a coalbed methane well and the effect of engineering development. Accurately predicting the daily gas production is a key issue to ensure the efficient production of coalbed methane. Taking the No.3 coal seam of Zhengzhuang block in southern Qinshui Basin as the research object, the dynamic parameters of CBM well drainage such as bottom hole flow pressure, liquid column height, casing pressure, daily water production and stroke frequency, were selected as independent variables to analyze the daily gas production. The correlation between these drainage parameters established a CBM well gas production prediction model and method based on cyclic neural network. The results show that, there is a positive correlation between the gas production of CBM well and stroke, gas production is negatively correlated with bottom hole flow pressure, casing pressure, liquid column height and daily water production. Based on feature importance analysis in deep learning random forest algorithm, the nonlinear relationship between the drainage and production parameters and daily gas production was studied and the contribution rate of drainage parameters to daily gas production in forecast model, the order of importance of the influence of drainage and production parameters on daily gas production was shown as follows: bottom hole flow pressure >liquid column height >casing pressure >daily water production > stroke frequency. On that basis, an improved prediction model of short and long time memory based on cyclic neural network, substitute the drainage and production data of Well Z4-026 into the model for calculation, the gas production of CBM Wells in the next 60 days was predicted, the prediction results were compared with the traditional support vector machine regression model, random forest regression model and BP neural network model, the prediction results are compared with the traditional support vector machine regression model, random forest regression model and BP neural network model. It is found that the improved shortand longterm memory neural network prediction model has relatively good fitting effect. The error between actual and forecast production is less than 5%. In the prediction and analysis of gas production of five CBM wells in Zhengzhuang block, the relative error is less than 10%. Therefore, this method will provide an effective way for the prediction of CBM gas production and the formulation of reasonable drainage and production system.

0 引言

1 LSTM算法原理与模型

   1.1 循环神经网络

   1.2 LSTM神经网络

2 模型参数优选

   2.1 日产气量与排采参数之间相关性分析

   2.2 基于相关系数的排采参数分析

   2.3 产气量影响因素分析

3 LSTM模型拟合

   3.1 数据选择和归一化

   3.2 模型训练

   3.3 模型预测对比与分析

4 结论