Accurate forecasting of labor productivity is key to enabling proactive project controls. Prior research has focused on forecasting weekly labor productivity trends using auto-regression (AR) models, modeling work-packages individually. Results are often limited to one-step-ahead forecasts, with models unable to automatically account for the factors that impact productivity.

This paper presents a case study on the implementation of a hybrid method based on AR and deep neural networks for time-series multi-step forecasting of daily labor productivity at the activity level. We demonstrate the forecasting capabilities of the model, which automatically accounts for the impact of factors on labor productivity trends.

In this paper, we present a novel hybrid method based on Auto-Regression (AR) and deep neural networks for time-series modeling of short-term labor productivity at the construction project level. Prior research on productivity forecasting had focused on AR models fitted with least-squares method. These classic AR methods are easily interpretable but they are slow to fit to the large volume of productivity data captured on jobsites. Also, specific assumptions are made on how resources or project constraints influence productivity rates and at best, these models can accurately predict one step ahead-of-time. Deep neural networks show promise in solving the scaleability problem; However, they can be complex for typical time-series data. They also lack interpretability which is key for labor productivity and root-cause analysis.

To bridge the gap between conventional AR methods and deep learning, we present ARRC-Net: Auto-Regression Resource-Constrained Network using a feed-forward neural network approach. Using real-world productivity data from three construction projects, we demonstrate that the ARRC-Net: (1) automatically learns sparse AR-coefficients while making it interpretable similar to classic AR methods; (2) has a linear computational complexity compared to conventional AR methods which are quadratic; (3) scales well to forecasting productivity at multiple steps ahead of time, making it ideal for look-ahead planning purposes; (4) eliminates the need for making explicit assumptions about the influence of project constraints on productivity.

The presented ARRC-Net predicts short-term productivity at both work-location and trade-level, continually and proactively.

Accurate forecasting of daily construction production trends requires complete and uninterrupted observations. However, the collection process of daily observations for each construction activity faces human resource constraints, resulting in incomplete collected production trends. To work with such data, most traditional imputation model techniques rely on linear interpolation, which introduces bias in the production trends, negatively affecting the predicting accuracy of any forecasting model.

To overcome this challenge, this paper explores the feasibility of using a deep neural network model architecture called Variational Autoencoders to learn latent features represented by the mean and variance of a complete time-series of construction production rates. These features are then used to handle missing data points of similar production time-series, resulting in accurate and complete production trends. A case study on the installation of drywall panel activities is conducted using data collected from the construction of three real multi-house family residential buildings. We demonstrate the potential of our deep learning method architecture in outperforming traditional imputation models.

Over the past few years, the construction industry has seen a surge in reality capture solutions. Camera-equipped drones and hardhat-mounted 360 cameras are commonly used to visually document the as-is conditions of project sites. This availability of reality capture data and the pressure for more transparency in project execution has created a demand for solutions that automatically track progress. Despite more than a decade of research, many practical and theoretical limitations prevent Scan-vs-BIM methods from scaling into real-world projects.

Using experimental data from one hundred real-world projects, this paper explores the impact of accuracy, completeness, and frequency of reality capture; Level-of-Development; granularity of BIM; as well as Scan-vs-BIM alignment on the state-of-the-art geometrical- and appearance-based progress monitoring solutions. The theoretical challenges with recognition of BIM objects and materials from images and a path-forward for using synthetic data for training deep learning methods for automated recognition are discussed in detail.