【作者】 金长宇；

【导师】 马震岳；

【作者基本信息】 大连理工大学 ， 水利水电工程， 2007， 博士

【摘要】 由于水电站地下厂房开挖跨度大，其构造和受力条件复杂，因此围岩稳定一直都是地下工程领域中的研究重点和难点。由于初始地应力场，岩体本构以及岩体参数都有很强的随机性、模糊性以及非线性，导致传统分析手段不能准确地对地下工程进行计算分析。多门学科交叉，多种手段并用是目前解决地下工程的有力武器。本文运用解析、数值以及智能方法从多个角度对水电站地下厂房的围岩稳定进行分析和评价，为设计和施工提供理论基础和可靠依据。全文的总体内容如下：（1） BP神经网络已广泛地应用于岩体力学参数和初始地应力场的反演分析，但在实际应用中，BP网络存在着网络训练易于过度、收敛速度慢、易陷入局部极小以及隐层节点数难以确定等缺点。基于上述原因，本文采用RBF网络作为反分析方法，利用有限差分格式的快速拉格朗日算法进行正分析计算，依据若干测点的应力数据，对计算区域的岩体力学参数以及初始应力场进行多参反演。以呼和浩特抽水蓄能电站的实例来说明RBF网络在反演分析的精度和学习速度方面，均优于传统BP网络的反演算法。（2）根据水电站地下厂房所在地域、地质条件以及工期的要求，采用何种开挖方式以及开挖高度，至今水工设计规范上尚没有明确的规定。本文以呼和浩特抽水蓄能电站快速施工方案设计为背景，利用FLAC^3D对5层和7层开挖方式进行敏感性分析，以求探讨减少开挖层数，加快施工进度的可行性及工程措施，并对两种开挖方案下的围岩稳定性进行评价，同时对开挖过程中的支护措施给出合理的参考意见。（3）基于锚杆的主动支护原理，将锚杆与岩体之间的剪应力转换为作用在洞室内表面上的面压力，将锚固后的岩体视为等效材料。采用Poynting-thomson流变本构模型，利用弹性力学和粘弹性力学推导得出锚固岩体的流变公式。以圆形引水隧洞为算例，应用FLAC^3D计算软件对其计算结果进行校核，通过比较可以发现，推导得到的公式不仅计算简单，而且其精度能够满足工程需要。（4）自适应网络模糊推理系统（ANFIS）是一种将神经元网络和模糊逻辑有机结合的新型模糊推理系统，采用反向传播算法和最小二乘法的混合算法分别调整前提参数和结论参数，这样充分地利用了神经网络和模糊逻辑的优良特性。本文将ANFIS和水电站地下厂房围岩变形的监测数据相结合，建立水电站围岩变形的预测模型。利用对现场观测数据的学习，对未来的围岩变形进行预测。通过工程实例，说明了该方法的合理性和可行性。（5）针对神经网络没有严格理论基础的缺陷，本文将一种改进的支持向量机—精确在线支持向量机（Accurate Online Support Vector Regression）方法应用到水电站地下厂房开挖过程的顶拱围岩变形预测中。以琅琊山抽水蓄能电站地下厂房顶拱变形为例，通过与现场观测值和其他预测方法结果的比较，说明精确在线支持向量机具有很强的学习能力和很高的预测精度。通过上述问题的研究，从多角度对围岩的稳定和变形进行了分析和评价，并且得到了有意义的结论和成果，最后本文给出了下一步研究工作的展望。更多还原

【Abstract】 Since excavation span of underground powerhouse is large, its geologic construct and state of stress is complex, the stabilization of underground powerhouse has always been the main focus in underground engineering. Because of initial stress field of rock masses, constitutive model and material parameters have randomness, fuzziness and strongly nonlinear, which result that underground engineering may not be accurately analyzed and simulated by traditional calculation method. However, along with the development of crossing multi-disciplinary, using diversiform methods are effective weapons to study the characteristic of underground engineering. The paper adopts the analytic solution, numerical methods and intelligent algorithms to analyze and evaluate the stabilization of surrounding rock of underground powerhouse and provides advices for the design and construction. The main contents are as follows:（1） At present, BP neural network has been widely used in back analysis of material parameters and initial stress field of rock masses in mechanics. However, BP neural network is prone to be over- trained, slow in convergence, not global minimum but local ones obtained and number of neurons in hidden layer hard to be determined. Due to above problems, the paper adopts RBF neural network and advanced BP neural network to identify mechanical parameters and initial geo-stresses according to actual observed normal stresses of some specific points. Direct computations based on fast Lagrangian analysis of Continuum （FLAC） are performed to get enough training samples for RBF neural network and BP neural network. The example of Hohhot pumped-storage power station shows that the combination of RBF neural network with FLAC is more effective and rapid than the application of BP neural network.（2） According to the requirement of terrain tract, geological conditions and project’s time limit of underground powerhouse, how to choose a better construction scheme and step has not always been determined by specification for the design of any hydraulic structure. On the base of the project background of Hohhot pumped-storage power station, the paper applies the FLAC^3D program to study the deformations, stresses and plastic states separately for five layers and seven layers of excavation height model. Comparing the results of two different models, the stability evaluation of its surrounding rocks has been advanced and some reasonable advices are gained in excavation and supporting.（3） Based on active timbering principle of bolt-grouting, shearing strength between bolt-grouting and rock can be transformed into the surface pressure of acting on the inner surface in the cavity and anchor-forced rock is regarded as a equivalent material. Adopting rheological constitutive relation model of Poynting-thomson, the rheological expressions of rock with anchoring force is advanced utilizing elasticity and visco-elasticity. Taking a circular tunnel for example, the precision of the simplified calculation results can satisfy the demand of actual project by checking a precedent with FLAC program.（4） Adaptive-Network-based Fuzzy Inference Systems （ANFIS） are new fuzzy inference systems which organically integrate with neural network and fuzzy logic. ANFIS adopt Hybrid learning algorithm of back propopagation algorithm and Least Square estimate method, which adjust premise parameter and conclusion parameter respectively, so as to make full use of excellent characteristic of neural network and fuzzy logic. This paper combines ANFIS and Hydropower Powerhouse surrounding rock deformation monitoring data to establish the forecasting model of Hydropower rock deformation. The future of rock deformation will be predicted by studying on-site observation data. The paper takes Longtan Power Station as engineering background to illustrate the rationality and feasibility of this method.（5） As neural network lacks strict theoretical foundation, the "Accurate Online Support Vector Regression" method is applied to study the evolution law for underground powerhouse rock surrounding’s stability. For example, the crown’s deformation of a pumped-storage power station, comparing with other prediction methods’ monitoring results; it shows that the learning efficiency and prediction accuracy of AOSVR is superior to other prediction methods obviously.In conclusion, the paper shows that surrounding rock’s stability and prediction of deformation have been studied and analyzed from different aspects. Finally a summary is given some problems to be further studied are discussed.更多还原