基于梯度提升决策树算法的膨润土膨胀力预测

Prediction of Bentonite Swelling Pressure Based on GBDT Algorithm

膨润土由于其高膨胀、低渗透和良好的核素吸附能力,被公认为高放废物深地质处置库的缓冲/回填屏障首选材料。实际工程中,高压实膨润土的膨胀力是缓冲屏障设计的关键指标之一,但影响因素众多、作用机制复杂,现有理论模型难以综合考虑多种因素的影响。为此,运用作为数据分析可靠手段的机器学习方法,采用泛化能力出众的梯度提升决策树(GBDT)算法,以目前世界上的典型膨润土类型为研究对象,基于对国内、外相关实验研究成果的搜集与梳理,搭建了基于GBDT回归算法的膨润土膨胀力预测模型,并结合既有实验数据进行了验证与应用效果分析。GBDT算法的预测准确率可达91.7%,明显优于随机森林等其他回归算法,模型可为我国已开建的地下实验室现场试验与处置库建设等提供技术参数和理论支撑。

Bentonite is the preferred buffer/backfill barrier material for deep geological disposal repository due to its high swelling ability,low permeability and high nuclide adsorption capacity.In actual engineering,the swelling pressure of highly compacted bentonite is one of the key indexes for buffer/backfill barrier design.Previous studies showed that there are many factors affecting the swelling pressure of bentonite and the mechanisms are complex,it is difficult for the existing theorical models to analyze the influence of multi-factors comprehensively.Therefore,one of the most effective machine learning algorithms,Gradient Boosting Decision Tree(GBDT),was used in this study to comprehensively analyze the influence of different parameters and predict the swelling pressure.Compared with other machine learning algorithms,the GBDT model showed much higher precision with a value of 91.7%.The GBDT regression model is expected to provide technical parameters and theoretical support for the field test of underground laboratory and the construction of disposal repository built in China.