基于优化随机森林的H2S腐蚀产物类型及腐蚀速率预测

目的研究H2S环境下碳钢腐蚀产物类型及失重腐蚀速率预测模型,为含硫油气田管道腐蚀防护设计与选材提供依据。方法整合H2S腐蚀模拟实验数据,采用随机森林算法对各腐蚀因素重要性进行排序,一方面以腐蚀产物类型为输出量,通过随机森林分类算法建立硫铁腐蚀产物类别预测模型,另一方面以腐蚀速率为输出量,通过随机森林回归算法建立腐蚀速率预测模型,并与其他模型进行比较。运用网格搜索方法对各类算法的超参数进行优选,以提高预测可靠性。结果随机森林算法得出的影响H2S腐蚀产物类型的因素重要性排序为:H2S分压、温度、pH值、实验周期、总压、CO2分压。基于网格搜索优化的随机森林分类模型交叉验证得分超过0.9,f1得分达到0.96,优于其他三种常用分类模型。采用网格搜索优化的随机森林回归模型预测结果与实际值的均方误差为0.86%。相关系数R值为0.979,优于其他两个回归模型。结论网格搜索优化后的随机森林分类、回归模型对含H2S复杂环境下的碳钢腐蚀产物类型及腐蚀速率预测准确性较高,能够为油气田管道腐蚀防护提供参考。

Prediction of H2S Corrosion Products and Corrosion Rate Based on Optimized Random Forest

The work aims to investigate the prediction model of iron sulfide corrosion products and corrosion rate under the H2S environment, and provide basis for designing and selecting the corrosion protection of pipelines in gas fields with sour oil. Based on the collected experimental data of H2S corrosion, the priority of each corrosion factor was ranked by the random forest algorithm. On the one hand, the prediction model of the corrosion product category was established by the random forest classification algorithm with the corrosion product types as output. On the other hand, the prediction model of the corrosion rate was established by random forest regression algorithm with corrosion rate as output. The grid search method was used to optimize the super parameters of various algorithms to improve the prediction performance. According to the random forest algorithm, the characteristic importance of H2S corrosion product types was ranked as follows: H2S partial pressure, temperature, pH, experimental period, total pressure and CO2 partial pressure. The cross-validation score of random forest classification model based on grid search optimization exceeded 0.9 and f1 score reached 0.96, which was better than other three common classification models. The mean square error between the prediction result and the actual value of the random forest regression model with grid search optimization was 0.86%. The R value of the correlation coefficient was 0.979, which was better than the other two regression models. The random forest classification and regression models optimized by grid search have high accuracy in predicting corrosion product types and corrosion rates in complex H2S environment, which can provide reference for the corrosion protection pipelines in oil and gas fields.