An innovative technique for improving productivity forecasting models

Productivity forecasting of construction operations has gained tremendous momentum in both the construction industry and academia. Many of the developed models utilize clustering methods in order to recognize existing hidden patterns among the historical data and improve the modelling performance by mimicking these patterns. It is shown that the improper determination of the number of clusters in such models can noticeably distort their fitness, which is not treated well in the literature. This paper scrutinizes the impacts and benefits of optimizing the number of clusters in productivity forecasting models. To this end, Subtractive Clustering is applied to optimize the clustering performed by K-Means method. A set of internal indices that consider Separation and Compactness of the resulted clusters are used to validate the method. The proposed technique is further investigated for a Neural-Network-Driven Fuzzy Reasoning (NNDFR) model developed to simulate a construction operation, in which several qualitative and quantitative factors are considered. Empirical results show that the model performance, in terms of Mean Squared Error (MSE), improves by up to 60 percent when the optimal number of clusters is determined using the presented technique. The developed technique benefits researchers and practitioners to improve the accuracy of modelling in productivity estimation based on a set of construction historical data.