基于人工智能的钻速预测模型数据有效性下限分析

钻速预测对于优化钻探工艺、降低作业成本、实现科学钻探具有重要意义,它是钻探钻井作业的一项重要内容。基于人工智能的钻速预测精度令人瞩目,但该技术需求的海量数据对传统钻探钻井作业的要求较高。为明确使用人工智能建立钻速预测模型的最少数据量,本文基于中国南海10口井的21917条数据进行了分析。通过相关性分析,所有的输入参数可被划分为高、中、低相关性3大类。通过逐步引入参数建立预测模型和对比预测精度,发现当引入的参数数量足够时,3种相关性参数均可建立起高精度(≥85%)的预测模型。引入参数的相关性越高,建立高精度预测模型需求的参数量越少。通过逐步扩大取样间隔的方式,对比发现所有的预测模型均呈现随取样间隔的增大、预测模型的准确性降低的规律。而预测模型建模的取样间隔下限可通过寻找精度降低时的拐点获得。经过验证,在数据维度与取样精度均为下限时,基于3种相关性参数建立的BP神经网络预测模型仍然能够获得较高的预测精度。

Discussion on the lower limit of data validity for ROP prediction based on artificial intelligence

Prediction of ROP is of great significance for optimizing drilling technology,reducing operating costs,and realizing scientific drilling,and it is also an important part of drilling operations.The accuracy of ROP prediction with artificial intelligence technology is remarkable,but the massive data required by this technology puts forward higher requirements for the traditional drilling operations.To determine the minimum amount of data for ROP modeling based on artificial intelligence,analysis was carried out based on 21917 data samples collected from 10 wells at South China Sea.Through correlation analysis,all input parameters were divided into three categories:high,medium and low correlation.By gradually introducing parameters to establish a prediction model to compare the accuracy,it was found that when the number of parameters was sufficient,the parameters in all three categories can be used to establish a high?precision(≥85%)prediction model;however,the higher the correlation of the parameters,the less the number of the parameters required to set up a high accuracy prediction model.When the sampling interval is gradually expanded,comparison found that the accuracy of all the prediction models decreased with the sampling interval increased.The lower limit of the data sampling interval for setting up the prediction model can be obtained through finding out the downward inflection point of prediction accuracy.It is verified that the BP neural network prediction model based on any of the three correlation parameters can still obtain high prediction accuracy when both data dimension and sampling accuracy are at the lower limits.