BP神经网络法预测顺北超深碳酸盐岩储层应力敏感程度

顺北油田是典型的深层应力敏感性油田，由于深井取心困难或者没有足够岩心支撑实验，同时岩样非均质性强，无法测试应力敏感程度，影响储层物性和产能评价的准确性。收集鹰山组7口井与应力敏感相关的测井、试井等6种测试数据，渗透率、裂缝宽度等7种参数，以及岩石组分、储层温压等资料，利用单相关分析和灰色关联分析筛选11个应力敏感伤害主控因素，加载至所建BP神经网络，设定激励函数和网络参数进行训练直至达到期望误差，在训练好的模型中加载已知应力敏感伤害结果的输入层参数进行计算，将计算结果与已知结果比较，应力敏感伤害程度预测符合率为100%，平均预测误差7.15%。再利用所建网络对顺北一间房组进行预测，并以室内实验值作为基准，计算得到应力敏感渗透率伤害率和临界应力预测误差均小于10%，表明该模型同样适用于其他碳酸盐岩储层应力敏感伤害预测。研究结果表明，BP神经网络法可以用来预测碳酸盐岩储层的应力敏感程度，能够解决深层碳酸盐岩储层由于取心困难导致无法支撑实验的难题。

Application of BP neural network method to predict the stress sensitivity of ultra deep carbonate reservoir in Shunbei Oilfield

Shunbei Oilfield is a typical stress-sensitivity deep oilfield. Coring in deep wells is difficult or there are not sufficient cores to support the experiments while the rock samples are of strong heterogeneity, so the stress sensitivity cannot be tested, so as to impact the evaluation accuracy of reservoir physical properties and productivity. In this paper, 6 kinds of testing data (e.g. logging and well testing) and 7 parameters (e.g. permeability and fracture width) related to stress sensitivity and rock composition and reservoir temperature and pressure data of 7 wells in the Yingshan Formation were collected. Then, 11 main control factors of stress-sensitivity damage were screened out by means of single correlation analysis and gray association analysis and loaded into the established BP neural network. Excitation function and network parameters were set for training until the expected error was reached. The input layer parameters with known stress-sensitivity damage result were loaded into the trained model to carry out calculation. The calculation results were compared with the known results. It is shown that the coincidence rate of predicted stress-sensitivity damage degree is 100% and the average prediction error is 7.15%. What’s more, the established network was used for the prediction of the Yijianfang Formation in Shunbei. Taking the laboratory test value as the reference, the calculated stress-sensitivity permeability damage rate and the error of predicted critical stress are both less than 10%, indicating that this model is also applicable to the stress-sensitivity damage prediction of other carbonate reservoirs. In conclusion, BP neural network method can be used to predict the stress sensitivity degree of carbonate reservoir and can solve the difficulty that the difficult coring in deep carbonate reservoirs fails to support experiments.