基于BP、RBF及GAAA-RBF神经网络的工程估价方法

发布时间：2018-05-29 18:58

本文选题：智能算法 + 工程估价　；参考：《华侨大学》2013年硕士论文

【摘要】：在建设项目前期，快速而准确的估算出工程项目的造价，对项目的投资决策具有十分重要的意义。随着智能算法的深入研究，工程估价领域出现了许多新的预测方法，如多元回归、神经网络、案例推理、遗传算法、小波分析等。论文提出一种利用蚁群算法对径向基神经网络的结构参数进行组合优化的改进智能算法，并将该算法用于建筑工程造价的快速估算。论文首先收集了70组中标工程数据，在充分理解和学习数据预处理技术的基础上，结合工程造价数据的具体特征，提出了包括数据清洗、数据转换和数据约简等内容的数据预处理方法和流程，得到55组可利用的工程案例库，为后续建模提供数据基础。其次，本文分别建立了基于改进BP神经网络工程估价模型、调用NEWRB函数的RBF神经网络和基于K-均值聚类的RBF神经网络两种工程估价模型，并对3种模型进行比对分析。分析结果表明：BP神经网络在构建相对困难，需要定义很多复杂的参数如学习速率、动量因子的大小、隐含层节点数目等，这些参数的设计并没有严格的理论指导，需要进行估计。RBF神经网络无论是采用何种构建方法，构建起来均相对简单，当调用NEWRB函数构建RBF时，其性能主要取决于分布宽度spread，而基于K均值聚类的RBF神经网络主要取决于重叠系数和隐含层数。整体来说,基于K均值聚类的RBF神经网络的学习速度更快、预测精度更高，但是学习效果有待提高。最后，为了能进一步提高基于K均值聚类的RBF神经网络的学习效果和预测精度，本文引进了遗传-蚁群混合算法对RBF的主要结构参数中心矢量、基宽向量和网络权重进行组合优化，建立基于改进蚁群算法的RBF神经网络工程估价模型。并结合厦门市中标工程数据进行学习仿真，，仿真结果表明，该通过参数优化的RBF神经网络工程造价预测模型，预测误差在5%以内，网络泛化能力更优越，可用于实际工程造价的辅助估算。
[Abstract]:In the early stage of the construction project, it is very important to estimate the cost of the project quickly and accurately, which is very important to the investment decision of the project. With the development of intelligent algorithms, there are many new prediction methods in the field of engineering evaluation, such as multivariate regression, neural network, case-based reasoning, genetic algorithm, wavelet analysis and so on. In this paper, an improved intelligent algorithm based on ant colony algorithm is proposed to optimize the structural parameters of radial basis function neural network, and the algorithm is used to estimate the construction cost quickly. Firstly, 70 groups of winning engineering data are collected. On the basis of fully understanding and studying the data preprocessing technology, combined with the specific characteristics of the engineering cost data, the data cleaning is proposed. The data preprocessing method and flow chart of data conversion and data reduction are used to obtain 55 groups of available engineering case bases, which provide the data basis for subsequent modeling. Secondly, two engineering evaluation models based on improved BP neural network, RBF neural network based on NEWRB function and RBF neural network based on K-means clustering are established, and the three models are compared and analyzed. The analysis results show that it is relatively difficult to construct the WBP neural network, and many complicated parameters such as learning rate, momentum factor size, number of hidden layer nodes, etc., need to be defined. The design of these parameters is not strictly guided by theory. No matter what method is used to construct the. RBF neural network, it is relatively simple to construct. When the NEWRB function is called to build the RBF, Its performance mainly depends on the spread of distribution width, while the RBF neural network based on K-means clustering mainly depends on the overlap coefficient and the number of hidden layers. On the whole, RBF neural network based on K-means clustering has faster learning speed and higher prediction accuracy, but the learning effect needs to be improved. Finally, in order to further improve the learning effect and prediction accuracy of RBF neural network based on K-means clustering, this paper introduces the genetic ant colony hybrid algorithm to the center vector of main structural parameters of RBF. The RBF neural network engineering evaluation model based on the improved ant colony algorithm is established by combining the base width vector and the network weight. Based on the engineering data of Xiamen City, the simulation results show that the prediction error is less than 5%, and the generalization ability of the network is better. It can be used to estimate the actual engineering cost.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP183;TU723.3

【相似文献】