School of Petroleum Engineering, China University of Petroleum (East China)
National Engineering Research Center for Oil & Gas Drilling and Completion Technology
Key Laboratory of Unconventional Oil & Gas Development, China University of Petroleum (East China), Ministry of Education
Dongbei Oil and Gas Branch, China Petroleum and Chemical Corporation
China United Coalbed Methane Co., Ltd.

针对钻井过程中储层裂隙发育情况不明，防漏堵漏方法和材料选择困难的问题，结合实际井史资料，提出了一种基于神经网络的储层裂缝宽度预测方法。首先通过相关性分析，对储层裂缝宽度的主要相关因素进行分析并对其进行排序，选取了泵压、钻井液排量以及钻速等7种主要相关因素作为输入参数并利用附加动量算法和变学习率算法对模型收敛速度进行提升，同时对模型结构进行优选。其次利用遗传算法(GA)和Adaboost算法对BP神经网络(BPNN)进行优化，克服了其易陷入局部极小值的问题，提升了模型的预测精度。最后建立了Adaboost-GA-BP神经网络预测模型对储层裂缝宽度进行预测研究，同时对比分析了模型的预测精度。结果表明，相关参数与储层裂缝宽度的相关性由高到低依次为漏失速度与漏失量、泵压、钻井液排量、钻速、井深、塑性黏度和钻井液静切力。另外，附加动量算法和变学习率算法使得训练结束时训练数据的绝对误差和降低了27%，显著提升了模型性能，同时通过GA算法优化模型的权值和阈值以及利用Adaboost算法进行集成优化进一步提升了预测精度，建立的Adaboost-GA-BP神经网络储层裂缝宽度预测模型误差和相关系数分别为18%和0.98，与随机森林等其他模型相比，模型的预测精度高，可为勘探开发过程中的裂缝宽度计算以及堵漏方案的制定提供一定的指导。

During the drilling, it is difficult to select the appropriate leakage prevention and plugging methods and materials as the development of reservoir fractures is unknown. Herein, a reservoir crack width prediction method based on neural network was proposed with reference to the actual history data of wells. Firstly, the main influencing factors of reservoir fracture width were explored and ranked by correlation analysis, and seven main influencing factors, including pump pressure, drilling fluid displacement and drilling speed, were selected as the input parameters. The rate of convergence of the model was improved using the additional momentum algorithm and the variable learning rate algorithm, and the model structure was optimized. Secondly, the Genetic Algorithm (GA) and Adaboost algorithm were used to optimize the BP neural network (BPNN), overcoming the problem of its tendency to fall into local minima and improving the prediction accuracy of the model. Finally, an Adaboost-GA-BP neural network prediction model was established to predict the reservoir crack width, with the prediction accuracy analyzed comparatively. The research results show that the parameters correlated with the reservoir crack width include the loss rate and loss amount, pump pressure, drilling fluid displacement, drilling speed, well depth, plastic viscosity and static shear force in a descending order. In addition, the additional momentum algorithm and the variable learning rate algorithm can reduce the sum of train set absolute error at the end of training by 27%, significantly improving the model performance. At the same time, the weight and threshold of the model were optimized by the GA algorithm, and the integrated optimization was realized by the Adaboost algorithm to further improve the prediction accuracy. Thus, the final reservoir crack width prediction model established based on Adaboost-GA-BP neural network has a root mean square error (RMSE) and a correlation coefficient of 18% and 0.98, respectively. Compared with other models such as the random forest, the model proposed herein has a higher accuracy, which could provide some guidance for the calculation of reservoir fracture crack and the preparation of leakage plugging scheme during the process of exploration and development.

中国石油集团公司关键核心技术攻关项目(2022ZG06)；国家科技重大专项(2016ZX05044-003-001)

王健，徐加放，赵密福，等. 基于神经网络的钻井液漏失裂缝宽度预测研究[J]. 煤田地质与勘探，2023，51(9)：81−88.

WANG Jian，XU Jiafang，ZHAO Mifu，et al. Prediction of crack width of drilling fluid leakage based on neural network[J]. Coal Geology & Exploration，2023，51(9)：81−88.