摘要
尾矿坝位移趋势表现为显著的动态特性,传统的静态位移预测模型难以反映尾矿坝动态位移趋势。针对上述问题,提出一种将粒子烟花算法(Particle Fireworks Algorithm,PFA)、动态窗口的在线学习(Dynamic Window-based Online Learning,DWOL)机制和门控循环单元(Gated Recurrent Unit,GRU)相结合的PFA-GRU-DWOL的尾矿坝位移实时预测方法。首先,为提升模型的计算精度和运算效率,提出了一种新型的PFA优化算法;其次,结合PFA和GRU,构建基于PFA-GRU的尾矿坝位移预测初始模型,并实时预测尾矿坝位移;再次,对比预测及实测结果,基于移动平均误差自动控制在线学习频率,构建了一种DWOL在线学习机制,在动态更新预测模型的基础上实现尾矿坝位移的实时预测;最后,以CEC2022基准函数集和攀西地区某尾矿坝位移预测为例,验证了算法及模型的有效性。结果表明:PFA算法在参与对比的7种算法中达到了最佳的收敛速度及稳定性;PFA-GRU-DWOL模型预测结果的M_(AE)、M_(APE)、R_(MSE)和R^(2)分别达到0.21、0.45%、0.28和0.99,明显优于对比模型。研究成果可为尾矿坝位移动态实时预测提供一种新的思路。
The displacement trends of tailings dams display notable dynamic characteristics,rendering traditional static prediction models inadequate for accurately capturing their behavior.To tackle this challenge,this study introduces a novel approach that integrates the Particle Fireworks Algorithm(PFA),a Dynamic Window-based Online Learning(DWOL)Mechanism,and Gated Recurrent Units(GRU)for real-time displacement prediction of tailings dams.This method is referred to as PFA GRU DWOL.Initially,this method leverages the strengths of Particle Swarm Optimization(PSO)and the Fireworks Algorithm(FWA)to create a new optimization algorithm known as Particle Fireworks Algorithm(PFA).This algorithm utilizes PSO for global exploration and FWA for localized dense search.Subsequently,PFA is employed to optimize the hyperparameters of the GRU network,resulting in the construction of an initial PFA GRU prediction model that effectively captures the spatiotemporal dependencies of tailings dam displacement.Thirdly,to address the limitations of traditional offline models in adapting to dynamic data,a DWOL mechanism has been developed.This mechanism dynamically updates training data using sliding window technology and adaptively adjusts the model's update frequency based on the Moving Average Error(MAE),ensuring real-time and accurate predictions.To evaluate the effectiveness of this approach,algorithm performance was compared using the CEC2022 benchmark functions,and empirical analysis was conducted with displacement monitoring data from tailings dams in the Panxi region.The experimental results demonstrate the following findings:(1)Compared with PSO,Genetic Algorithm(GA),Moss Growth Optimization(MGO),Fata Morgana Algorithm(FATA),Harris Hawks Optimization(HHO),and Polaris Leader Optimization(PLO),the PFA exhibits superior convergence speed and stability.(2)The PFA GRU DWOL model significantly outperforms both the static GRU and GRU DWOL models,achieving evaluation metrics of M_(AE)=0.21,M_(APE)=0.45%,R_(MSE)=0.28,and R^(2)=0.99,indicating a high level of prediction accuracy.This study offers a novel dynamic real-time prediction method for the safety monitoring of tailings dams while also providing valuable technical insights applicable to displacement prediction in other geotechnical engineering contexts,such as slope stability and dam safety assessments.
作者
唐宇峰
何俚秋
胡光忠
TANG Yufeng;HE Liqiu;HU Guangzhong(School of Mechanical Engineering,Sichuan University of Science&Engineering,Yibin 644005,Sichuan,China)
出处
《安全与环境学报》
北大核心
2026年第2期529-536,共8页
Journal of Safety and Environment
基金
四川省高校重点试验室开放基金项目(2024WYJ01)
泸州市重点实验室项目(SCHYZSA-2024-01,SCHYZSB-2024-01)
四川轻化工大学“652”科研创新团队项目(SUSE652A004)。
关键词
安全工程
粒子烟花算法
门控循环单元
在线学习
尾矿坝位移
实时预测
safety engineering
particle fireworks algorithm
gated recurrent unit
online learning
tailings dam displacement
realtime prediction
