Modeling and predicting the variation of US highway construction cost

Deep learning;Ensemble learning;Highway construction cost

Cao, Yang ; Ashuri, Baabak Building Construction Tien, Iris Irizarry, Javier Song, Xinyi Shahandashti, Mohsen Li, Shuai ; Ashuri, Baabak

University:Georgia Institute of Technology

Department:Building Construction

关键词: Deep learning;    Ensemble learning;    Highway construction cost;   

Others  :  https://smartech.gatech.edu/bitstream/1853/61237/1/CAO-DISSERTATION-2019.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

The U.S. government attaches great importance to highway construction every year. Because of the importance of highway construction projects and the tremendous expenditure, the budgeting and cost control is a significant job for all federal agencies and state Department of Transportation (DOTs). A major problem for highway construction costs is that they exhibit a significant variation over time such as the highway construction cost indexes (HCCI), and it is mainly caused by a complex interactive effect such as market-related and project-specific factors. The variation hinders the estimators to catch the correct trend of the market and thus poses a challenge for both owners, such as state, local Department of Transportation, and contractors, in correct budgeting and cost estimating. The main objective of this PhD research is to explore the explanatory variables and develop machine learning methods to model and predict highway construction cost and examine the accuracy of the forecasting results with those of the existing methods such as regression analysis and time series models. The major promising feature of the proposed methods over existing ones is the capability to explain the non-linear relation, which is prevailing in practice. Besides, there is no restriction for the proposed models, compared to linear regression, which assumes the linear relation and normal-distributed-error, and time series analysis, which requires the stationarity of data before analysis.The dissertation summarizes the results from two parts of the research: (1) Univariate time series index prediction in Long Short-Term Memory; (2) Modeling the unit price bids submitted for major asphalt line items in Georgia, project-specific and macroeconomic factors. A deep learning model based on the recurrent neural network will be developed due to its strength in long term memory and catching the variation of the data. The Texas HCCI is used as an illustrative example because Texas reports a frequent volatile index. The performance of the model will be tested in different prediction scenarios (short-term, midterm and long-term). A dataset containing fifty-seven variables with potential power to explain the variability of the submitted unit price bid are collected in this study. These variables represent a wide range of factors in six aspects: project characteristics, project location and its distance to major supply sources for critical materials, level of activities in the local highway construction market, overall construction market conditions, macroeconomic conditions, and oil market conditions. The machine learning feature selection algorithm Boruta analysis will be used to select the most significant features with the greatest capability to predict the unit price bid for asphalt line items. The partial dependence plots endow the explanatory power to the developed machine learning models. An ensemble learning model will be constructed based on the selected features to forecast the unit price bid. The accuracy of the predicted machine learning models will be compared and validated with the existing multiple regression model and the Monte Carlo Simulation. The main contribution of this research to the body of knowledge in cost forecasting are be summarized from three aspects: first, the research developed a new set of machine learning models that provide more accurate costs forecasts compared to the existing methods. For example, the non-linear machine learning methods are more accurate than the time series models which are frequently used in the former research. In the field of cost research, a small improvement in model accuracy results in a significant amount of actual impact in budget estimation. Second, the modified encoder and decoder architecture performed well in numerical time series data prediction problem. Instead of making one sequence of output, the roll-forward forecasting turned out to be more accurate. Third, from the practical application perspective, the proposed machine learning models can handle a wide range of issues with the input data that are common in the field of highway construction cost forecasting, such as missing values. Another practicality contribution is that methods in this research are applicable to big data which is an industry trend, while most former models were developed based on a small dataset. The research also proposed a construction cost database which will largely provide the convenience for easy utilization of the model. Fourth, the research identified the most significant features to forecast the variation of unit price bids of resurfacing projects in Georgia, and the analysis laid emphasis on the explanatory power of prediction models. With the improved prediction capability, state DOTs can benefit from the proposed models in preparing more accurate budgets and cost estimates for highway construction projects. The analysis process and proposed models in research are also applicable to other time series data prediction and cost estimating problems.