储层预测的代价敏感主动学习算法

Reservoir prediction through cost⁃sensitive active learning

传统的储层预测需要耗费大量的时间且对研究人员的专业能力要求极高,采用人工智能方法实现储层预测可以有效地改善预测效率.然而,因为环境、设备等原因导致油气井数据中存在大量属性值缺失,大大降低了储层识别精度.针对属性值缺失造成分类困难的问题,提出一个统一评估和动态选择的代价敏感主动学习算法(Active Learning Algorithm with Unified Evaluation and Dynamic Selection,ALES):(1)考虑各种代价的设置和计算,包括误分类代价、属性代价、标签代价和样本代价;(2)使用softmax回归实现对属性值和标签价值的统一评估;(3)提出一种具有排列组合和贪婪策略的最优获取方案,实现属性值和标签的动态选择.采用三个真实测井数据进行实验,显著性实验分析证明了ALES的有效性及其相对于监督代价敏感分类算法和缺失填补算法的优越性.

For oil and gas industry,traditional reservoir prediction usually takes a lot of time and requires researchers to have high expertise,while using artificial intelligence to realize reservoir prediction effectively improves the efficiency of prediction.However,due to environmental and equipment reasons,there are a large number of missing attribute values in oil and gas well data,which greatly reduce the accuracy of reservoir identification.To solve the problem of classification difficulty due to the lack of attribute values,we propose a cost‐sensitive active learning algorithm with unified evaluation and dynamic selection(ALES).First,we consider the setting and calculation of various costs,including misclassification costs,attribute costs,label costs and sample costs.Second,we use softmax regression to achieve a unified evaluation of attribute values and label values.Third,we propose an optimal acquisition scheme with permutation and greedy strategies to achieve dynamic selection of attribute values and labels.The experiments used three actual logging interpretation data.The results of significance test verify the effectiveness of ALES and its superiority to the state‐of‐the‐art supervised cost‐sensitive classification algorithms and missing filling algorithms.