【作者】 彭光金；

【导师】 俞集辉；

【作者基本信息】 重庆大学 ， 电气工程， 2010， 博士

【摘要】 随着市场经济的发展,尤其是招投标制度的推行,传统定额概预算制度已经无法完全满足工程建设的需要,有必要引进在国外广泛应用历史数据估算工程造价的做法来解决这一问题。由于工程项目建设的历史数据收集比较困难,具有数量小、属性多的特点,基于历史数据的造价估算实质上是小样本数据的学习问题,比海量数据的学习更困难,不能采用基于样本无穷大假设的经典统计方法进行处理。近二十年来,模糊数学、灰色关联度以及神经网络等理论技术应用于造价估算的研究较多,但是应用模糊数学和灰色关联度理论设计的算法和模型过于简单,神经网络学习算法的收敛性、鲁棒性以及泛化性较差,都还无法满足实际应用的要求。支持向量机的出现为小样本学习问题的解决提供了最佳的理论技术平台,论文融合粒子群算法、聚类技术等人工智能技术,对支持向量机进行改进研究,论文研究的目的是提出一种基于参数优化回归支持向量机的小样本数据智能学习改进算法,同时把该算法应用于工程造价快速估算。论文首先对小样本数据预处理技术进行研究,结合工程造价历史数据的具体特点,提出包括数据清洗、数据转换和数据约简等内容的小样本数据预处理方法,并且以电力输电工程为案例进行仿真,验证数据预处理方法的有效性。其次,在粒子群算法引入排斥速度和时变领域等概念,提出自适应多种群粒子群优化改进算法(Self Adopt Multiple Particle Swarm Optimization,简称SAMPSO),给出改进算法的数学模型和工作流程,仿真表明改进算法可以解决多峰函数优化问题,在多峰函数寻优时既可以全局寻优,也可以找到所有局部最优点,多峰函数寻优效率更高。然后,应用粒子群改进算法对小样本数据的分割聚类进行优化,提出一种基于SAMPSO的两阶段聚类改进算法,给出改进算法的数学模型和基本工作流程,并且以电力输电工程为案例,对改进算法进行仿真,与模糊聚类算法和粒子群优化算法的聚类效果进行比较,仿真结果表明,该算法分类效果明显优于模糊聚类算法等普通聚类算法。接下来,应用SAMPSO算法、遗传算法对支持向量机的参数寻优进行了对比分析,仿真结果表明,基于SAMPSO算法的参数寻优效果更好,应用SAMPSO算法对小样本数据的非线性核主元分析( Kernel Principal Component Analysis,简称KPCA)进行优化,基于参数寻优的支持向量机和非线性主元分析方法,提出参数寻优的支持向量机智能学习改进算法(Prameters Optimization Support Vector Machine Algorithm,简称POSVMA),给出其数学模型和工作流程,并且以电力输电工程为案例,对改进算法进行仿真,和标准支持向量机算法的仿真预测结果进行比较分析,验证改进算法的合理性和有效性。最后应用上述理论分析研究成果,基于POSVMA改进算法,提出一种基于小样本历史数据的工程造价快速估算方法,给出其具体工作流程,以电力输电工程和变电工程为案例进行仿真,仿真结果表明,该造价快速估算方法基本能够满足工程造价管理和控制的实际应用需要。更多还原

【Abstract】 With the development of market economy, especially implementation of bidding system, traditional budget making system mainly based on quotas can not meet the need of project construction, in order to solve such question, it is necessary to introduce estimation method of project cost based on history project datas which is widely used in foreign contries. Because history datas are difficult to be collected, in most case, they are small samples, so cost estimation based on history project datas is essentially the question of small samples learning which is more difficult than massive datas learning, classical statistical methods based on hyposesis of infinite samples can not be used to solve such question. In the recent twenty years, theories such as fuzzy mathematics, grey correlation and neural network were widely used for research on cost estimation, but algorithms and models according to fuzzy mathematics and grey correlation theory were too simple, meawhile convergence,robust and generalization ability of neural network theory was poor, they can not meet the need of practical application. Support vector machine (SVM) technique is the best choice for research on small samples learning. In this paper, other artificial intelligence theories such as particle swarm optimization algorithm and clustering technique are used to improve regression SVM , a kind of intellegent learning modified algorithms of small sample based on parameter optimization regression SVM is proposed, meanwhile such modified learning algorithms is used for rapid estimation of project cost.Firstly, per-processing technique of small sample data is studied, combined with specific characteristics of history project data, a kind of specific small sample data per-processing method which include data cleaning, data conversion and data reduction is proposed, moreover, an illustrative example of transimission line projects proving the efficiency of the method is given. Secondly, new concepts named repulsive velocity and auto-tuning territory are introduced to improve particle swarm optimization algorithm, Self-adaptive multi-grouped PSO (abbr. SAMPSO) is proposed, its mathematical model and work flow are given, which successfully solve optimization problem of multi-modal function, optimization efficiency of SAMPSO is higher than standard PSO for multi-modal function, specially all local optimal points of multi-modal functions can be founded out by SAMPSO. Thirdly, SAMPSO is used to improve the performance of clustering of small sample data, two stages clustering modified algorithm is proposed, its mathematical model , work flow and an illustrative example of transimission line projects proving the efficiency of modified algorithm are given, Simulation results show that efficiency of such modified algorithm is better than conventional clustering algorithm such as FCM. New modified algorithm powerfully improve the performance of SVM clustering and regression on convergence , robust and exactness. Fourthly, SAMPSO and GA algorithms are used for parameter optimization of SVM at the same time, simulation results show that SAMPSO has better performance than GA, meanwhile nonlinear kernel principal component analysis (abbr. KPCA) is used to process small sample data, based on SVM and KPCA, a regression SVM modified algorithm(Prameters Optimization Support Vector Machine Algorithm, abbr. POSVMA) is proposed, its mathematical model and work flow are given, simulation results of transimission line projects by POSVM and SVM are compared to verify the efficiency of modified algorithm . At last, using above research achievements, based on POSVMA, a rapid estimation method of project cost is presented, its work flow is given. Simulation results of transimission line projects and power transformation projects show that such method can basically meet the practical need of management and control of project cost.更多还原