Machine Learning-Based Surrogate Models for Construction Project Duration Prediction

This paper presents a novel multivariate regression and optimal artificial neural network (ANN) model that is capable of precisely predicting construction project completion duration time. To achieve this, real-world data on the construction of residential units, industrial buildings, and large apartments were garnered through standard publications and reports rolled out by construction companies for devising the models. A database of 193 project instances was considered for model development. The inputs for the models are, the number of total activities, activities that were considered to be critical during the planning stage, the number of slack activities, maximum and minimum duration consumed by a critical activity (in days), and the output is the project duration. Among the five regression models developed, the cubic polynomial fared well with a high R2 value of 0.725, and the ANN model of topology 5-8-9-8-1 emerged to be optimal with a high R2 value of 0.93. Levenberg Marquardt (LM), conjugate gradient descent (CGD), and Quasi-Newton (QN) were also tested for optimal training. CGD algorithm was found to be the best one with low Akaike's information criterion value (AIC). The major takeaway from this work is avoiding the huge drudgery involved with conventional methods like CPM and PERT.