A Deep Learning Approach for Forecasting Construction Project Duration at Completion

Abstract

Accurate forecasting of project duration is crucial during the execution phase as it affects its overall performance, timely decision-making, identification of potential delays, and resource allocation. This research proposes a proof of concept based on artificial intelligence, specifically using deep learning algorithms, demonstrating its potential application. These algorithms have shown remarkable results in finding patterns in large amounts of data and making accurate predictions. Moreover, the dataset provided to the model is treated as a time series, capturing the sequential nature of data collected throughout the execution phase. Additionally, predictions are yielded at work package level, providing to project managers granular information to make decisions. The study follows these steps: (1) a comprehensive literature review was conducted to explore the latest advancements on related topics and underline current gaps. (2) A data acquisition model was elaborated founded on a consistent selection of duration-influencing factors. Then, actual data was collected and profiled from multiple projects using their work package names as links, thus creating datasets per work package. (3) The forecasting duration at completion model was developed, including data preprocessing and the computational deep-learning-based modelling per work package. After that, the overall project duration was modelled using the Critical Path Method and Precedence Diagramming Method and set in the Graphical User Interface. (4) The developed forecasting model was used to comparing three well-suited deep-learning algorithms with actual project data and consequently, selecting the most accurate. Next, the selected algorithm was incorporated into the Graphical User Interface. The study also was validated by comparing the proposed model with traditional methods, including the Earned Value Methodology (EVM) and Earned Schedule Methodology (ESM). Finally, the resultant model was verified through sensitivity analysis. As a result, the forecasting model based on the Long Short-Term Memory (LSTM) algorithm demonstrated the best performance against Multi-Layer Perceptron and Convolutional Neural Network algorithms. Likewise, it performed better than intensive-used forecasting methods in the industry, such as Earned Value Methodology (EVM) and Earned Schedule Methodology (ESM). These promising results contributes to the foundation of Artificial Intelligence (AI) applicability in construction project duration at completion forecasting.