Full Automated AEC Scheduling Framework using Machine Learning

Abstract

The AEC industry is unfortunately one of the slowest to embrace technological advancements. And therefore, the managerial aspect of something like planning and scheduling is very reliant on heuristic processes and decisions by people in command with little use of previous historical data. Historical data accessibility is a non-issue, manually mining relevant information for future projects, on the other hand is just too timeconsuming due to unique intricacies of projects, and differences in construction methodologies. This master’s thesis aims to propose a synergistic framework for the automation of AEC scheduling using Machine Learning through holistically studying all accessible previous projects’ schedules to provide useful insights, feedback, and in due course automate a large portion of baseline schedules in the process of being generated for new projects.

This entailed the development of optimized models for studying different aspects of scheduling. These aspects are most probable task list generation, activity relationships, and optimized durations. There is more than one approach provided for the activity relationships and task list generation models. Each of the alternative approaches are evaluated for suitability. And a final cohesive Inference File was created to use all the models for a final output.

Alternative approaches for activity relationships are utilizing a GNN and utilizing a FineTuned LLM, while the final model utilizes a traditional classification layout with a Dataset format that involves more effort to develop. The first model for task generation is utilizing LLM approach for the work was already done on it from the previous model; in addition to a Sequence-to-Sequence LSTM RNN with a horizontal and vertical task list formatting. Finally, the durations model utilized Sci-Kit Learn’s library for regression analysis with multiple model comparisons whereby the best model was selected according to a weighted average criterion set by the author. The datasets used to train, and validate these models were a fully synthetic built upon augmentation, sensitivity analysis, and randomness of a real-world dataset. Training and validation scores of each model indicate a promising ability to automate the scheduling portion significantly. Additionally, effort was put to ensure that the whole framework flows together.