摘要
复合驱由于强碱化学剂的注入易造成严重的结垢现象,而预测清防垢加药量是制定科学的清防垢方案的关键。为了获得一种具有较强实用价值和较高预测精度的复合驱清防垢加药量的预测方法,该文将随机森林回归算法应用到复合驱清防垢加药量预测。首先,分析并处理影响清防垢加药量的20维特征,利用随机森林回归的特征重要性评估功能对影响加药量的众多特征进行筛选;然后,通过网格搜索和K折交叉验证的方式得到最优参数组合,建立随机森林回归的复合驱清防垢加药量预测模型;最后,通过大庆市某采油厂三元复合驱某区块的清防垢数据,验证预测模型的有效性,并与CART回归、SVM回归和ANN进行对比。实验结果证明,该方法可以实现清防垢加药量的有效预测,较对比模型的预测精度高出约23%,且其预测效果较为稳定,同时经过特征筛选的方式可以提高模型的预测精度,约比未经特征筛选的随机森林回归模型高1.86%。
Due to the injection of strong alkali chemical agent,the compound driving is easy to cause serious scaling phenomenon,and predicting scale cleaner and scale inhibitor dosage is the key to formulate a scientific scale cleaner and scale inhibitor scheme.In order to obtain a more practical value and a more accurate prediction method of compound drive’s scale cleaner and scale inhibitor dosage,we try to apply random forest regression to predict compound drive’s scale cleaner and scale inhibitor dosage.First of all,we analyze and deal with 20 characteristic factors that affect the amount of scale cleaner and scale inhibitor dosage,through the characteristics importance evaluation function of the random forest regression to conduct feature screening for numerous features affecting the dosage.And then,we determine the optimal model parameters by grid search and cross validation and establish the prediction model of compound drive’s scale cleaner and scale inhibitor dosage based on random forest regression.Finally,we collect the historical data of ternary compound drive’s scale cleaner and scale inhibitor dosage in Daqing city,analyze the validity of the model experiments,and make a comparison with decision regression tree,support vector machine and artificial neural network.Experiment shows that the proposed prediction method can effectively predict the scale cleaner and scale inhibitor dosage,at least 23%better than the average prediction accuracy of other models,and the prediction effect is stable.At the same time,the method of feature screening can improve the prediction accuracy of the model,which is about 1.86%higher than the random forest regression model without feature screening.
作者
李春生
张圣权
张岩
张可佳
LI Chun-sheng;ZHANG Sheng-quan;ZHANG Yan;ZHANG Ke-jia(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)
出处
《计算机技术与发展》
2022年第1期204-209,共6页
Computer Technology and Development
基金
国家自然科学基金(51774090)
黑龙江省自然科学基金面上项目(F2015020)
东北石油大学青年科学基金(QJ-007-002)
东北石油大学研究生教育创新基金(JYCX_CX07_2018_1)。
关键词
复合驱清防垢
加药量
随机森林
特征筛选
网格搜索
交叉验证
compound drive’s scale cleaner and scale inhibitor
dosage
random forest
feature selection
grid search
cross validation