RandWPSO-LSSVM反演方法及其在大型地下工程中的应用

聂卫平^{1, 2}，徐卫亚³，王伟³

(1. 广东省电力设计研究院，广东 广州，510663；2. 清华大学 土木水利学院，北京，100084；3. 河海大学 岩土工程科学研究所，江苏 南京，210098)

摘 要：根据粒子群算法(PSO)和支持向量机(SVM)的特点和局限性，对其进行改进，建立随机权重粒子群-最小二乘支持向量机(RandWPSO-LSSVM)反演模型，由正交设计、均匀设计和三维有限元数值计算给出学习和测试样本，反演糯扎渡水电站大型调压井工程区的围岩力学参数和初始地应力场。研究表明：RandWPSO对LSSVM预测模型的优化效果明显比PSO好；反演得到工程区x向和y向侧压力系数分别为2.641 5和2.083 1，微新岩体、弱风化下层、弱风化上层、全强风化层岩体、断层等围岩的弹性模量分别为24.849 2，10.898 7，2.839 8， 0.270 4和 0.651 3 GPa；反馈计算得到的位移相对实测位移误差较小，验证反演参数的合理性，也表明所建立的RandWPSO-LSSVM反演模型合理可靠，可有效指导大型地下工程参数设计和施工稳定性分析。

关键词：随机权重粒子群；最小二乘支持向量机；反演；大型地下洞室；弹性模量；调压井

RandWPSO-LSSVM inversion method and its application in large-scale underground engineering

       NIE Weiping^{1, 2}, XU Weiya³, WANG Wei³

(1. Guangdong Electric Power Design Institute, Guangzhou 510663, China;2. College of Civil and Hydraulic Engineering, Tsinghua University, Beijing 100084, China;3. Institute of Geotechnical Engineering, Hohai University, Nanjing 210098, China)

Abstract:The PSO and SVM were improved because of their characteristics and limitations, and the RandWPSO-LSSVM inversion model was established. After the learning and test samples were obtained by orthogonal experimental design, uniform experimental design and three-dimensional finite-element computation, the mechanics parameter of adjoining rock and initial stress field of large-scale surge shaft engineering of Nuozhadu hydro-power station were inversed. The results show that the optimization effect of RandWPSO is better than that of PSO for the LSSVM prediction model. Through inversing, x direction lateral pressure coefficient is 2.641 5 and y direction lateral pressure coefficient is 2.083 1, and the modulus of the fresh rock mass, the upper bed of slightly weathered rock mass, the lower course of slightly weathered rock mass, the strong weathered rock mass and the fault are 24.849 2, 10.898 7, 2.839 8, 0.270 4 and 0.651 3 GPa, respectively. The error between displacement obtained by feedback calculation and displacement obtained by measurement is little, therefore, the parameter obtained by inversion and the RandWPSO-LSSVM inversion model are established reasonably, and the parameter design and stability analysis in construction of large-scale underground engineering could be guided effectively.

Key words:random weight particle swarm optimization; least squares support vector machine; inversion; large-scale underground cavern; modulus; surge shaft