Shale Gas Production Forecasting Using Additional Weights LSTM Neural Network—A Case Study of the HS Well Area

现阶段机器学习作为页岩气产量预测的一种有效手段，在数据规模受限的开发新区却往往无法进行准确预测。针对此问题，提出了附加权重的LSTM神经网络预测应用。首先使用HS页岩气井区地质参数、工程参数和生产参数的训练数据建立标准LSTM预测模型，获得产量预测拟合公式；其次通过机器学习方法交叉分析确定影响因素对测试产量的影响权重，然后在初始LSTM模型引入附加权重来优化预测矩阵并进行产量预测，最后应用模型对区块老井与新井的各项参数进行优化。预测结果表明，附加权重后的LSTM模型预测值与真实值吻合度更高，优化结果可靠，可以很好地解决数据量受限井区的机器学习分析与优化问题。Machine learning serves as an effective means for shale gas production forecasting at the current stage. However, accurate predictions are often challenging in newly developed areas with limited data. To address this issue, an additional weights Long Short-Term Memory (LSTM) neural network prediction approach is proposed. Initially, a standard LSTM prediction model is established using a training dataset comprising geological parameters, engineering parameters, and production parameters from the HS shale gas well area, obtaining a production prediction fitting formula. Subsequently, a machine learning approach is employed for cross-analysis to determine the impact weights of influencing factors on test production. These weights are then introduced into the initial LSTM model as additional weights to optimize the prediction matrix and enhance production forecasting. The results indicate that the LSTM model with additional weights yields predictions that closely align with real values, yielding reliable optimization results and providing an effective solution to machine-learning analysis and optimization in data-scarce well blocks.