基于Stacking集成学习的超声粒径分布反演方法研究OA

The intricate and computationally-demanding constraints associated with conventional inversion models used for analyzing sus-pended sediment particle size distribution are delved into.A pioneering approach is presented which leverages the combination of base learners of extreme random trees,decision trees,and XGBoost,along with the meta-learner of random forest algorithm,denoting a new stacking ensemble learning model.The methodology encompasses the preparation of suspended sediment samples with predetermined size distributions achieved through sieving techniques,supplemented by ultrasonic attenuation experiments to establish the dataset.The process involves feature selection using the Boruta algorithm and fine-tuning model hyperparameters through an improved grey wolf opti-mization algorithm.In contrast to the individual,yet high efficient random forest algorithm model,the proposed model showcases remark-able accuracy improvements,leading to notable reductions in relative root mean square error:4.644 7%for RR distribution,0.207%for log-normal distribution,0.381%for beta distribution,and 0.397%for random distribution.Furthermore,the model demonstrates robust generalizability during assessments involving bimodal,geometric,and U-shaped distributions,thereby offering a reliable methodology for inferring the particle size distribution of suspended sediments.