基于机器学习的页岩气采收率预测方法

Shale gas is a kind of unconventional natural gas existing in mud shale, which consists of adsorbed gas, free gas and solution gas. The proved reserves of natural gas from shale in China are abundant and widely distributed, which are usually buried deeper than 3,000 meters. The key technologies of shale gas production are horizontal well and hydraulic fracturing, while a difficult challenge for efficient production is to predict the recovery. If the recovery can be predicted, the degree of reservoir reconstruction can be evaluated, and the effects of the current construction parameters on gas production can be determined. In recent years, with the rise of deep learning, it has become a popular method for solving engineering problems. In this paper, we summarize the microseismic and fracturing construction parameters from field data. The data of horizontal wells in shale gas production in Fuling District, Sichuan Province are analyzed, and the categorical variables are processed with one-hot encoding, through which the features of fracturing are described by other continuous variables. The basic conception of three machine learning methods, SVR, DNN and XGBoost, is introduced, and the specific parameters for establishing the prediction models from reservoir and construction parameters to the recovery factor are illustrated. The convergence of models is stated by the decline of cost function and the increase of accuracy rate with the training process. The results show that, with a small sample size, the XGBoost model has higher forecast accuracy than the SVR and DNN models. Considering the difficulty of obtaining field data and long sampling period in fracturing, the XGBoost and other tree models are primary choices which can improve the forecast accuracy and stability. The advantages and disadvantages of various models are analyzed, and the importance of relevant parameters is also discussed. Through the whole procedure of establishing a prediction model, how to take diverse types of features into account, describe them in proper data forms and appropriate models contribute directly to the predicting performance. In summary, this paper sets up reasonable recovery prediction models under small data condition in shale gas construction, which provid the guidance for adjusting construction parameters and can be applied to predict shale gas production.