Martin Fischer, Maximilian Schuetz
In the field of construction scheduling research has recently shown that it is possible to automatically generate close-to-optimum construction schedules (Morkos, Dong, et al.) and to dramatically reduce the time for such a schedule to be created.
While these schedules are a perfect plan to start with, many studies have also shown that there is often significant variability (Garcia, Schuetz) in the actual execution of the plan over the course of a project.
With our research we therefore intent to create a system that supports the construction project team to focus on the right aspect of the right tasks at the right time in order to stay close to the optimum schedule.
We initially intended to create a system that learns from historic schedule data to predict the reliability of a task. After reviewing several schedule archives of numerous general contractors, we concluded that the data was neither of the right quality, nor consistently collected in order to apply Machine Learning methods. We therefore focused our efforts on creating a tool that collects the right information in the right way, yet without any additional effort compared to current practices.
We studied 14 construction projects around the world and found that they all planned in similar cycles (Phase Plan, Monthly Look-Ahead, Weekly Work Plan, and Daily Check-in). We did not, however, observe that results of one planning cycle were used to adjust the next planning cycle. Most projects tracked metrics, such as Percent Plan Complete (PPC), but they did not use them to change the planning method for the next cycle. The Look-Ahead windows of each planning cycle were rigid and stayed the same (e.g. 2 months), regardless of the project’s performance or the task’s characteristics. A task for steel columns, for example, had a longer lead time for ordering material, precast concrete beams were highly depended on shop drawings 4 months before on-site delivery, whereas drywall tasks could be made ready on shorter notice. Yet most projects strictly looked ahead for 2 months during their monthly Make-Ready meetings and did not examine the status for material orders, equipment orders, or shop drawings for tasks beyond this Look-Ahead window.
As a core element to establish feedback loops across planning cycles, we must gather as much data about tasks and their reliability as possible. One source we are going to examine is historic schedule data of the past years. Since our first tests, however, showed that historic schedule data might not have the required quality, nor consistency, to formally learn from, we are going to integrate two more sources of input.
For once, we plan to integrate the experience of senior project stakeholders. This should serve as a knowledge management foundation for a possible prototype.
Further we plan to systematically collect data with a new prototype in the structured way, which we could not find in historic schedule archives so far.
In a common Lean practice, the team should be able to discuss Takt Area by Takt Area, with the 3D viewer adapting to their actions. Each task needs to be examined for its Make-Ready categories and tasks missing critical requirements are flagged by the system.
In order to isolate separate supply chain categories, you should be able to filter tasks by Make-Ready categories, either showing which tasks and objects are still missing shop drawings or which processes already have workspace available right now.
During a Weekly Work Plan meeting, the team can discuss week by week and commit to upcoming tasks. If a task, for example in week 3, cannot be executed, because say there are still no drawings available, the team could then call up a backlog of similar activities in the same Takt Area that are ready and can be executed instead.
On a Daily Planning level, the team can again review today’s task’s in the morning or at the beginning of a shift and log the task status as “doing”. At the end of the day or a shift, the task status can then be switched to either complete or incomplete.
It is very important to us, that the system can learn as much as possible about each task, without the user specifically entering that information. So with the connection to 3D objects, we do not only know the task’s description, company, and reliability, but can also map data from the BIM like quantity, quality specs or assembly method.
During our studies we mapped these data and besides typical metrics like overall reliability and root cause analysis, we were able to create certain risk profiles by BIM object types like columns, beams, slabs or risk profiles by assembly methods i.e. how reliable is cast-in-place in comparison to pre-cast. We could also compare the reliability of a task to the actual lead time of the approval for its shop drawings and therefore calculate an optimum lead time for shop drawings, for the task to be on time.
All these calculations, however, were done under great additional effort by our research team and merely highlight examples of what can be done.
If we want to make the method more reliable, we need to capture much more data. And since we found that there is very little data out there, that is already structured in this manner, it becomes vital to develop a prototype as the one we propose, and automate the process in the background.
With this approach of establishing these kinds of Feedback Loops in a BIM-based and Lean-enabled environment, we believe we can dramatically improve the reliability of monthly Make-Ready meetings and keep the schedule continuously close to optimum.
We have created a cloud-based web platform that accommodates all four major planning cycles that we observed during our initial case studies. The platform features a BIM viewer as well as a task viewer for Takt Time Planning. An algorithm in the backend can recognize Takt Areas in an imported BIM and automatically generate a Takt Plan in the plan viewer which is linked to the BIM. If the user group creates a task within this Takt Area it can also link it to the respective 3D object(s).
The tool is also capable of dynamically adapting the length of the Look-Ahead window during each planning session and individually highlights the tasks that the team should focus on during the current meeting.
The algorithm that is used to decide which task is to be highlighted at any given time can be taught corporate (or project) standards e.g. with regard to lead times for tasks under certain conditions. The tool also gathers a large amount of data points on every action taken in the system, which can be used to (1) improve the existing algorithm and (2) automatically generate customized performance reports for the project team.
Prototype tests on construction projects
In order to gather feedback on the performance of our tool under real world conditions we have tested the prototype on three different construction projects within the United States. The group of test projects was located in the State of California and consisted of a (1) Sports Arena project (2) residential building project and (3) large medical facility project.
All projects leveraged BIM methods and already had a set of tools and methods for construction planning for most Planning Cycles in place. In order to measure the impact of our tool on the projects we first analyzed the current state of practice and, if possible, measured the time effort of each planning process. Afterwards we implemented our tool to gather the same performance data with the same personnel on the project.
As a quantitative result we were able to reduce the effort of the planning sessions (the ones we were able to measure) by 58% from 5 h 30 mins per week to 2 h 20 mins per week. As a qualitative result we were able to help the client understand the implications of a change he had made and that had resulted in a change order request. After examining the condition with our tool the client agreed to the change order requested by the General Contractor.
The implementation of our tool on these projects has revealed that there is a significant potential for improvement in the area of construction planning across multiple planning cycles and that with tools like the one we developed we can help project teams to improve their planning performance. We have also learned, however, that requirements from each company (or even project) are different and that our tool will need to be enhanced with more features and metrics in order to serve the needs of most of the projects.