# Cost Estimation Model (Cem) for Residential Building using Artificial Neural Network

DOI : 10.17577/IJERTV5IS010431

Text Only Version

#### Cost Estimation Model (Cem) for Residential Building using Artificial Neural Network

Post Graduate Student, Civil Engineering Dept.,

Bhopal

Monica Vyas

Professor,

Oriental Institute of Science and Technology, Bhopal

Sanket Agrawal

Assistant Professor, Parul Institute of Technology,

Abstract The achievement of any project undertaking is defined by improved quantity and cost estimation technique that facilitates optimum utilization of resources. The objective of this study is to develop a cost estimation technique by using an artificial neural network (ANN) model that will be able to forecast the total structural cost of residential buildings by considering various parameters. In this study, data of last twenty three years has been collected from Schedule of rate book (SOR) and general studies. Eight input parameters, namely, cost of cement, sand, steel, aggregates, mason, skilled worker, non-skilled worker and the contractor per square feet construction were selected. The parameters were simulated in NEURO XL Version 2.1 for developing ANN architecture. The resulting ANN model reasonably predicted the total structural cost of building projects with correlation factor R-0.9960 and RSquared-0.9905 giving favorable training and testing phase outcomes.

Keywords Artificial Neural Network (ANN), Correlation Factor, Cost estimation, Model, Variables

1. INTRODUCTION

Cost is probably the first to be considered when it comes to construction projects. Accurate estimation of quantities and costs incurred in a construction project is a crucial factor in its achievement [1]. Because of the complexity of the construction industry and individuality of every project undertaking, several factors may affect the overall project cost. A number of objects, such as the structural, architectural, sanitary, electrical and air-conditioning system workings conclude the total cost of buildings. Olotuah, 2002, observed that building resources incur approximately 60% of the total cost of a residential building [2]. Meanwhile, the structural casing covers 25% of total construction cost in a multi-storey reinforced concrete residential building. Therefore, greatest worry must be exercised in the design of structural systems if a considerable reduction in cost is preferred. However, consistency and accuracy of available cost estimating techniques are matter of concerns presently. Thus, now there is a growing need to deal with the concerns by introducing a new and alternative approach for estimation of cost and to identify the factors responsible for variation of cost.

Usually, the ordinary least squares regression approach is applied and the replica is selected based on the coefficient of determine, R2. However, because of high correlation among a great group of variables, this technique tends to generate regression coefficients estimators that will badly perform in the presence of multi-collinearity [3]. Furthermore, the variance of the usual least squares estimator become inflated, which results in the low prospect of the estimator being close to the correct value of the regression coefficient [4]. This can be rectified by determining uncorrelated variables to be included in the regression model. Lots of quantity and cost estimation models have been developed. Linear regression is a very useful statistical device for analyzing and predicting the input of a potential new item to the overall approximation. Several other methods have been applied for the cost estimation, such as, principle component study, case- based reasoning and ANN.

Artificial neural networks can model complex non-linear relationships and approximate any assessable function. ANN is a powerful means to handle non-linear problems and subsequently map relations between complex input/output data and address uncertainties [5]. They are particularly helpful in problems where there is a complex relationship between an input and output. The main advantage ANN models have over physically based models is that they are data-driven and underlying contact using examples of the desired input output mapping [6].

Ismaail EISawy et al (2011) tried to develop a parametric cost-estimating model fifty-two actual real-life cases of building projects, in Egypt, constructed during 2002-2009 and achieved an accuracy of 80% [7]. Jamshid Sodikov (2005) developed a more accurate estimation technique for highway projects in developing countries at the theoretical phase using artificial neural network [8]. H.Muarat et al (2004) in Turkey, used training and testing data from thirty projects of 48 storey reinforced concrete structure by neural network methodology, achieving an average cost estimation accuracy of 93% [9]. Emad Elbeltagi (2014) developed an ANN model to predict the cost of highway structure projects in Libya by considering various factors that influence the highway construction [10].

The motive of this work is to explore the ANN technique and predict the total structural cost of buildings and to determine the factors which affect the cost of buildings. Hence a cost estimation model (CEM) has been developed using artificial neural networks, particularly multi-layer feed forward neural networks. The back propagation knowledge algorithm is used to instruct the network by iteratively processing a set of training sample and compare the network's prediction with the real. The variation in the estimation is propagated to the input for adjusting the coefficients. To accomplish this goal, structural cost data from past twenty three years from 1993 to 2015 were collected and were used simulated in ANN SOFTWARE.

2. METHODOLOGY

In this study a frame work has been developed for estimating the variation & construction cost for residential building. All the models where developed using data of input variables like cost of cement, sand, steel, aggregate, mason, skilled, non-skilled, Formwork, Brick work etc (Table 1). Also the rates of material that is input variables has been taken from Schedule of rate book of last 20 years. Pentium 4 class P.C. with window XP operating system is used to run NEURO XL Version 2.1 Artificial Neural Network software. For all the models, total data set is divided into training set (68%), validation set (16%), and test set (16%).

Table 1 Input parameters for ANN model.

 Year Cement (Rs/bag) Sand (Rs/ft3) Steel (Rs/kg) Aggregate (Rs/ft3) Mason (Rs/day) Skilled Labor (Rs/day) 1993 80 2 6 2 110 50 1994 90 2 7 2 120 55 1995 100 3 9 2 125 60 1996 110 4 10 3 130 65 1997 120 5 12 3 135 65 1998 125 5 13 4 70 1999 130 6 15 4 145 75 2000 135 7 17 5 150 75 2001 135 9 17 5 155 80 2002 140 12 18 6 155 90 2003 145 15 19 6 160 95 2004 145 18 20 7 165 95 2005 150 19 21 7 170 105 2006 155 20 22 8 175 100 2007 160 22 23 8 180 120 2008 178 23 25 9 200 120 2009 190 24 28 10 220 150 2010 200 25 30 12 250 200 2011 200 25 33 16 300 250 2012 225 30 38 18 350 250 2013 260 40 40 20 400 300 2014 260 40 48 40 425 340 2015 290 60 34 23 450 360

Selection of parsimonious ANN structure was accomplished by first fixing the number of hidden layers and then choosing the number of nodes in each of these layers. For larger networks, computational costs are high and might overfit the training data with too many nodes. The presentation of one- hidden layer ANN is found to be better than two hidden layers ANN [11]. Nodes in the hidden layers, are very important for characteristic extraction from the patterns of input time series, they normally have a very small weight changes and learn very slowly [12, 13]. Here single

hidden layer is used to generate feed-forward entirely- connected neural network (multi-layer perceptron). For both hidden and output layer hyperbolic tangent activation function is employed. This function has a sigmoid curve and is calculated by using the following formula as in equation 1.1

F(x) = (ex – e-x) / (ex + e-x) —————- (Eq. 1.1)

The selected output range is [-1 to 1]. Sum-of-Square is the error function used to rate the quality of the neural network. Training is stopped when the number of iteration become 500. As iteration is a single complete presentation of the training set to the neural network. This is the simplest and the most commonly used condition to stop training. When a neural network begins to over-train (i.e. to memorize data instead of generalizing and encoding data relationships), the validation errors increase while training errors might still reduce in that case, one has to retrain and restore best network. When the training is completed the testing operation was performed. The performance of each network model was then evaluated by computing the mean absolute error for each model.

3. RESULTS AND DISCUSSION

The table 1.2 shows the actual and forecasted structural cost obtained through the ANN model. The output obtained from ANN model also includes Absolute and Relative errors in prediction. From the error percentage it is clear that maximum error is 8.58%, for year 2008, which is less than 10%. Hence the result indicates good estimation values with prediction above 90%.

Table 2 Output parameters obtained through ANN model.

 Year Contractor (Rs/ft2) Forecast Abs. Error Rel. Error Estimate 1993 290 312.40957 22.409572 7.73% Good 1994 320 325.48742 5.4874173 1.71% Good 1995 350 355.31718 5.3171772 1.52% Good 1996 370 378.10013 8.1001263 2.19% Good 1997 400 409.25739 9.2573891 2.31% Good 1998 425 425.69218 0.6921805 0.16% Good 1999 450 460.90113 10.90113 2.42% Good 2000 475 490.3637 15.363698 3.23% Good 2001 500 500.56361 0.5636101 0.11% Good 2002 525 550.93636 25.936358 4.94% Good 2003 550 591.69713 41.697126 7.58% Good 2004 600 614.38795 14.387947 2.40% Good 2005 650 655.11369 5.1136869 0.79% Good 2006 700 668.14794 -31.85206 4.55% Good 2007 750 735.36585 -14.63415 1.95% Good 2008 850 777.0985 -72.9015 8.58% Good 2009 900 881.23432 -18.76568 2.09% Good 2010 1000 980.20803 -19.79197 1.98% Good 2011 1070 1028.7587 -41.24135 3.85% Good 2012 1100 1102.7111 2.7111141 0.25% Good 2013 1200 1238.3095 38.309463 3.19% Good 2014 1300 1314.485 14.485034 1.11% Good 2015 1400 1303.8944 -96.1053 6.86% Good

The results were then plotted on the scatter graph between actual and forecasted values of last 23 years. The forecasted values generated through the ANN model were found to have very high correlation coefficient. The estimated correlation coefficient and R-squared coefficient for the results obtained were 0.9960 and 0.9905 respectively as shown in Fig. 1.

Fig. 1 The scatter graph between actual and forecasted values.

The Table 3 shows the summary of the co-relation factor between actual and predicted structural cost. The average Absolute Error (AE) obtained for Training Set and Test Set were 21.436 and 27.183 respectively. The predictions for each year from 1993 to 2015 indicated Good Forecasts for both Training (100%) and Test Set (100%), where the Tolerance level was 10% for Training Set.

Table 3 Summary of the results.

 Training Set Test Set No of Rows 19 4 Average AE 21.436 27.183 Average MSE 932.1315 1477.6754 Tolerance 10% 30% No of Good forecasts 19 (100%) 4 (100%) No of Bad forecasts 0 (0%) 0 (0%)
4. CONCLUSION

The motive of this work was to explore the ANN technique and predict the total structural cost of buildings and to determine the factors which affect the cost of buildings. The developed cost estimation model (CEM) with back propagation knowledge algorithm was used to instruct the network by iteratively processing a set of training sample and compare the network's prediction with the real. The generated model fairly forecasted the structural cost, where the correlation coefficient and R-squared coefficient were found to be 0.9960 and 0.9905 respectively. The average absolute error of training set 21.43 and that of test set was 27.18, with the error varying from 8.58% (maximum) to 0.11% (minimum), indicating good error deviation during training. The accurate conversion of practical field data into real time data can bring major change in the construction industry by forecasting the cost of any project. Artificial neural networks can model complex non-linear relationships and approximate any assessable function. The advance prediction of overall residential building cost can help the user in decisive planning. This model can also be used in future by various stakeholders, to study variation in the project cost, if the cost of various important resources like steel, cement, labor, etc. is changed.

REFERENCES

1. Roxas, Cheryl Lyne C., An Artificial Neural Network Approach to Structural Cost Estimation of Building Projects in the Philippines De La Salle University, Manila, Issue-8 March 2014.

2. Olotuah, A., Recourse to Earth for Low-Cost Housing Nigeria, pp.37, 123129, Issue- 2002.

3. Tewari, D. and Singh, K., Principles of Microeconomics, New Delhi, New Age International Publications, pp.342, 1996.

4. Montgomery, D. and Peck, E., Introduction to Linear Regression Analysis, New York, John Wiley and Sons,1982.

5. Alqahtani Ayedh and Whyte Andrew, Artificial Neural Networks Incorporating Cost Significant Items towards Enhancing Estimation for (life-cycle)Costing of Construction Projects Curtin University, Australia ,pp. 51-64 , Issue-2013.

6. Vyas Monika, Performance Forecasting Of Common Effluent Treatment Plant Parameters By Artificial Neural Network , ARPN Journal of Engineering and Applied Sciences, VOL. 6, Issue-1 January 2011.

7. ElSawy Ismaail ,A Neural Network Model for Construction Projects Site Overhead Cost Estimating in Egypt Thebes Higher Institute of Engineering Corniche Maadi , Egypt , Vol. 8 , Issue-3 May 2011.

8. Sodikov Jamshid, Cost Estimation of Highway Projects in Developing Countries Artificial Neural Network Approach Saitama University , Japan , Vol. 6, pp. 1036 1047 , Issue-2005.

9. Gunaydn H. Murat, A neural network approach for early cost estimation of structural systems of buildings Izmir Institute of Technology , Turkey , Issue-2 April 2004.

10. Elbeltagi Emad, Conceptual Cost Estimate of Libyan Highway Projects Using Artificial Neural Network Mansoura University , Cairo

, Vol. 4, pp.56-66 , Issue-8 August 2014.

11. Vyas Monika, Forecasting of Water Quality Variables using ANN Framework , Radharaman Engineering College, Bhopal.

12. Maier, H.R. and G.C. Dandy, The effect of internal parameters and geometry on the performance of back propagation neural networks: An empirical study. Environ. Model. Software, 13: 193-209, 1998.

13. Morcous G., Comparing the use of Artificial Neural Networks and Case-Based Reasoning in Modeling Bridge Deterioration McGill University , Canada , Issue-5-8 june2002.