Application of Neural Network in the Cost Estimation of Highway Engineering

Transcription

1 1762 JOURNAL OF COMPUTERS, VOL. 5, NO. 11, NOVEMBER 2010 Application of Neural Network in the Cost Estiation of Highway Engineering WANG Xin-Zheng School of Civil Engineering,Nanyang Noral University,Nanyang city P.R.China, DUAN Xiao-chen and LU Jing-yan School of Econoics and Manageent,Shijiazhuang Railway nstituteshijiazhuang city,p.r.china, Abstract Based on the BP neural network, this paper sets up the odel of cost estiation of highway engineering. The BP neural network odel is trained by a saple data obtained fro soe perfored typical engineering to coe true quick cost-estiating.t is sure that the ethod is practical and the estiating results are reliable according to lots of exaples.t shows the proising perspective of BP Neural Network in cost estiate of construction engineering. ndex Ters -neural network, highway engineering, cost estiation. NTRODUCTON The cost estiation of the project is an iportant content of the feasibility Study. The accuracy of the cost estiation directly affects the project s decision, construction s scale, design schee, econoical effects, and project s proceeding. Estiating the project handily, quickly and exactly has the great significance for the anageent and control of the project s estiation [1]. The ain characteristic of the project s cost estiation is that there are too any factors that can affect the fabrication cost of the project, but little available related aterial[2]. The fabrication cost is affected by not only the construction s characteristic, but also any uncertain factors. And there is highly nonlinear apping relationship between the project s fabrication cost and the uncertain engineering characteristics. Therefore, how to reflect this apping relationship is the key of setting up the estiation odel. The relationship between these two ones is linearity in the traditional estiating ethods (such as regression analysis), which would reduce the precision and accuracy of the results. f the theory of neural network is introduced in the traditional ethod (Engineering coparison), it can reedy the shortage by the training process and ake the evaluation ore dependable and quicker.. METHODOLOGY OF NEURAL NETWORK The neural network is a new ethod of inforation processing [3]. t is the coplex network syste that is fored by any plentiful and siple processing units (neurons) [4]. The neural network is one kind of large-scale parallel connection echanis with adaptive odeling function, which siulates the structure of huan brain [5]. ts establishent is on the basis of odern neuroscience research. n any kinds of neural network odels, the back-propagation neural network odel (i.e. BP network odel) is the ost popular network because of its better functions of self-study and self-association. The standard BP network is coposed of three kinds of neurons layer. The lowest layer is called the input layer. [6] The iddle one is naed as the hidden layer (can be ulti- layer). And the top one is called the output layer. Every layer of neurons fors fully-connection, and the neurons in each layer have no connection. The learning course of BP algorith is coposed of two processes, propagation and antipropagation. n the propagation course, the input inforation is transferred and processed through input layer and hidden layer [7]. The state of every neural unit layer only affects the state of next layer. f the expected inforation cannot be got in the output layer, the course will turn into the antipropagation and return the error signal along the forer connection path. Via altering the value of concatenation weight between each stratu, the error signal is transitted orderly into the input layer, and then be sent into the propagation course. The repeated application of these two courses akes the error ore and uch saller, until it eets the requireents. The specific structure is as shown in figure 1: 2010 ACADEMY PUBLSHER doi: /jcp

2 JOURNAL OF COMPUTERS, VOL. 5, NO. 11, NOVEMBER Error back-propagation - + Desired output input layer hidden layer Output layer figuar1. Scheatic representation the basic structure of BP neural network n this figure, the relationship between the input and output of neural unit (except the input layer) is nonlinear apping, and S (Sigoid) function is always be adopted x to reflect it. f ( x) = 1 (1 + e ) is the Expression of output node function, and the differential coefficient is ' f ( x) = f ( x)(1 f ( x)).ts advantage is that the input data of any for can be transfored into the nubers that are in(0,+1).. APPLCATON OF NEURAL NETWORK N THE ESTMATON OF PROJECT S NVESTMENT The basic principle of the estiation and analysis of project s investent is that its establishent should be on the base of the siilarity of projects. For planned construction projects waiting for estiation, firstly, we start with the analysis of the construction s type and the project s feature, and find soe projects that are the siilar ones like the planned construction projects in any other projects that have been copleted. Then the fabrication cost aterials of these siilar projects can be used as the original data to ake the illation. At last, get the investent estiation and any other related data of the planned construction projects. When we use neural network to estiate the investent of project,we should analyze and settle the existing estiation aterials, analysis aterials and features of the forer typical projects in the light of the definite forats, and ake the as the training saples in order to input the into neural network to be trained, then finish the apping fro input layer (features of project) to output layer (estiation aterial). This apping is set up on the base of the siple nonlinear functions and expresses t coplex phenoenon of project s investent estiation. The neural network odel autoatically extracts the inforation and stores it as network weight in the inside of neural network. n this way, engineers and technicians can get the estiation of the project s investent by inputting the features of the planned construction projects and soe related price aterials into neural network. This paper akes the estiation of highway project s investent as the exaple and sets up an estiation odel of BP neural network by collecting the data of highway projects in soe area that have been copleted between 2000 and 2002 the data of sixteen typical projects were selected as training saples, thereinto the data of fifteenth and sixteenth are applied for checking.the estiation odel that is based on neural network is expressed in figure two. The odel can be divided into three sections: input-preprocessing odule, neural network odule, and output-processing odule. Neural network odule is the nuclear odule input-preprocessing odule neural network odule output-processing odule figure. cost estiation odel based on neural network The task of input-preprocessing odule is to preprocess the input data by changing qualitative stuff into quantitative ones so as to ake the calculation of neural network convenient the output-processing odule can change the output of neural network into the data of estiation that we need. A. quantitative description of engineering characteristics Engineering characteristics is the iportant factor that can express the project s characteristic and can reflect the ain constitution of the project s cost. The selection of the project s feature should refer to the statistics of historical projects aterials and expert's experience. Firstly, analyze the effect the typical highway project cost and the change of construction s paraeter ake to the estiation of the project s investent. we confir nine ain factors including landfor, highway grade, cross-section type (cutting, Ebankent, half-digging and half-filling), height, width, foundation treatent type, 2010 ACADEMY PUBLSHER

3 1764 JOURNAL OF COMPUTERS, VOL. 5, NO. 11, NOVEMBER 2010 aterial and thickness of the road surface, guard project type and so on as the project s features. And then List the characteristics of different engineering categories, each kiloeter highway engineering construction cost change by reason of construction cost influence's relevance according to quota and the project characteristic, ake copositor of the cost change and assigns corresponding quantification data subjectively, shown in table 1. TABLE. QUANTFCATON DATA OF CHARACTERSTCS OF ENGNEERNG CATEGORES quantitative value landfor aintain area hill plain type of foundation s cross-section cutting ebankent half-digging and half-filling highway grade high speed rank 1 rank 2 height of foundation s cross-section/ 0~ ~1 1~1.5 width of foundation s cross-section/ 0~10 10~15 15~20 type of foundation Ordinary replaceent plastic drain board consolidation geogrid aterial of road surface s structure asphalt concrete ceent concrete guard project coon guard anchor plates slope-protecting gravity retaining wall thickness of road surface s structure/ 0~ ~ ~0.4 quantitative value landfor the type of foundation s cutting half-digging and ebankent half-digging cutting ebankent cross-section half-filling and half-filling highway grade rank 3 the height of foundation s cross-section/ 1.5~2 2~2.5 ore than 2.5 the width of foundation s cross-section/ 20~25 25~30 ore than 30 the type of foundation sand pile drain consolidation dynaic copaction,ixing pile geotextile the aterial of road surface s structure guard project shotcrete wire support board girder support vegetation support the thickness of road surface s structure/ 0.4~ ~0.6 ore than 0.6 According to table 1, any highway project odel can be given a quantitative description. Taking Ti = ( ti 1, ti2,, ti9) T as an exaple, i is the serial nuber of project i( i = 1,2, ) t ( j = 1,2,,9) ; ij is the quantitative nueric value of the project i s feature j. For instance, soe highway project (the serial nuber is assued as i ) is in the plain, and the type of cross section is ebankent and highway grade for at a high speed with 1.8 height of roadbed cross section, 35 width and geogrid, asphalt concrete, coon guard and 0.45 thickness of road surface s structure. Therefore, its quantitative description is T i = (3,2,1,4,6,3,1,1,4). f soe feature is coposed of several kinds, count its weighed average calculated according to the proportion can be used as its quantification result. B. establishent of estiation odel of BP neural network This odel adopts three layers of BP network odel, and chooses the node-outputting function. there are nine units in output layer,which stand for project s characteristic vectors, such as landfor, highway grade, cross-section s type, cross-section s height, cross-section s width, foundation processing type, aterial of road surface s structure, road surface s thickness and protection type, the nine unite are arked as 1 ~ 9 ; the output unit is the estiation fabrication cost of highway project and expressed by 0. The nuber of hidden layer units is nineteen according to 2010 ACADEMY PUBLSHER

4 JOURNAL OF COMPUTERS, VOL. 5, NO. 11, NOVEMBER kologorov theore (2*9+1=19). Therefore, there are one hundred and ninety connections totally (9*19+1*19=190). The initial weight are chosen randoly fro rando nuber between(-1,1).the fixed network weight is the initial weight that has best training result. The selection of the initial weight has iportant ipaction the training result, because even the tiny change of the initial weight can lead draatic changes of error. According to the coplexity degree for the apping of inputting-outputting, fourteen training saples and twelve testing saples have been collect. Quantitative data and budget aterial of the sixteen typical saples are Listed in table 2 No. 1 TABLE. 2 3 QUANTTATVE DATA AND BUDGET MATERAL OF THE SXTEEN TYPCAL SAMPLES 4 nput Output(ten thousand yuan/k) O C. analysis of test results The results of group fifteen and sixteen tested the convergence network are respectively 12,500,000 yuan and 5,440,000 yuan. And the relative error between the real value and forecasting value is less than 5%. t can be seen fro the test s result that the overall error ratio is sall and the need for the estiation of engineering feasibility study can be basically satisfied. t shows that the odel has good generalization ability and the estiation odel is successful. V. CONCLUSON The neural network has got ore and ore attention in the econoic owing to its non-linear apping ability and approaching ability for any function. This paper uses the artificial neural network to extract the relation between the project s features and the estiation of fabrication cost fro the large nuber of past estiation aterials and sets up the estiation s neural network odel. The two test prove that the estiation accuracy eet the requireents, so this is an effective and feasible ethod that using neural network to estiate the highway project investent. Adopting neural network odel to estiate the highway project investent is copletely feasible and it can get the precise result. t has very iportant reference value and acadeic significance for Adopting new scientific ethod and prooting construction projects estiated investent research.. ACKNOWLEDGEMENTS The authors are grateful to the anonyous referees for their valuable rearks and helpful suggestions, which have significantly iproved the paper. This research is supported by the Soft-Science Progra of Henan Province (Grant No ); Natural Science Basic Research projects of the Education Departent of Henan Province (Grant No. 2009B630006); Soft-Science Progra of Nanyang City (Grant No.2008RK015) REFERENCES [1]. Zhou Liping, Hu Zhenfeng. The application of neural network in the cost estiation of construction [J].Journal 2010 ACADEMY PUBLSHER

5 1766 JOURNAL OF COMPUTERS, VOL. 5, NO. 11, NOVEMBER 2010 of Xi'an University of Architecture &technology, 2005,37(2): [2]. Duan Xiaochen, Yu Jianxing, Zhang Jianlong. A Method of Estiating WLC of Scheduled Railway Projects Based on CS,WLC and BPNN Theores[J].Journal of the China Railway Society, 2006,28(6): [3]. Han Yanfeng, Duan Xiangqian. Application of artificial neural network in date ining [J].Journal of Xi'an University of Architecture &technology, 2005,37(1): [4]. M. Kashaninejada, A.A. Dehghanib, M. Kashiri. Modeling of wheat soaking using two artificial neural networks (MLP and RBF)[J].Journal of Food Engineering, 2009,91: [5]. Federico Marini. Artificial neural networks in foodstuff analyses:trends and perspectives A review[j].analytica Chiica Acta, 2009,635(2): [6]. Hsiao-Tien Pao. Forecasting electricity arket pricing using artificial neural networks[j].energy Conversion and Manageent, 2007,48(3): [7]. Zhao Feng. Project nvestent Risk Evaluation based on BP Neural Network Syste[J].Construction Econoy, 2006,288(10):62-64 Xinzheng WANG, born in Nanyang, Henan Province, P.R.China,on Oct 23,1979.He received the aster degree in engineering fro Shijiazhuang Railway nstitute, China.Currently he is a lecturer in school of civil engineering, Nanyang Noral University, China. His current research interests are in the areas of cost estiation, engineering anageent. E-ail: wxz791023@126.co Xiaochen DUAN,born in Zhaoyuan, Shandong Province, P.R.China, on Jan.12,1962. He received the PhD degree fro Tianjin University, China, in Currently he is professor in Shijiazhuang Railway nstitute, His current research interests are in the areas of cost estiation, engineering anageent. He has published ore than 20 papers and 2 books. E-ail:zhangxp @163.co Jingyan LU,born in Baoding, Hebei Province, P.R.China,on Jan.28, 1980.she received the PhD degree in anageent fro Tianjin University, China, in Currently she is a lecturer in Shijiazhuang Railway nstitute. Her current research interests are in the areas of education, arketing, engineering optiization, logistics and supply chain optiization. E-ail:liujingyan999@163.co 2010 ACADEMY PUBLSHER