Artificial Neural Network Analysis for Cost Estimation of Building Projects in India

Abstract

In Construction Management, it is difficult to predict the cost estimate during the preliminary stage of the project because of limited information and unknown factors. Artificial Neural Networks can help in the prediction of estimate because of their simplicity and adaptability to non-linear problems. Due to their self-organizing nature they can be used to solve the problems even with low level programming. This makes them useful in interpreting and generalizing inadequate input information. ANN’s are crud e derivatives of the biological neural network with single layered or multi-layered neuron in the form of input layer, hidden layer and output layer. The neural network first has to undergo training from historical data in order to make predictions or show results. The size of the data set, number of hidden neurons and the neural network architecture determines the success of the results. Selecting the right data set becomes imperative in this case. For the purpose of cost estimation, the cost drivers were taken as inputs and their estimated costs were taken as the target value. The cost drivers were selected carefully through literature review and survey to provide more accurate results for the estimate. The main drivers identified were: type of building, location, seismic zone, project complexity, ground condition, soil condition, plot area, plinth area, built-up area, number of stories, number of basement, principal structural material, type of foundation, level of design complexity, modular design, market conditions, construction conditions, risk factor, impact of risk, estimated duration of work, specification, quality of work and detailed cost estimate of project. A Pareto analysis performed on the significant drivers showed that the Duration of Work, Complexity of the Building, Plinth Area and Built-up Area, Height and Specifications were the most important cost drivers in a construction project. A problem was formulated based on these drivers with numerical and categorical data. The data set was trained with a neural network using the MATLAB software using feed forward backpropagation. Training was carried out till the greatest correlation and least Mean Squared Error was obtained after multiple iterations. This trained data was used to predict the cost for a new project. The output of the testing was 87% accurate despite the small data set used.