Navigating multi-dimensional results from large parametric building simulation studies

Advances in computing in recent years allow for many thousands of building energy simulations to be computed in the time previously required for a single simulation run. Software tools exist that allow for a single input file to be modified in a number of different ways to generate thousands of self-similar input files which can then be automatically simulated. The problem with this approach is not the simulation time but the time and effort required for the analysis of the vast set of results generated. Large, multi-dimensional result sets cannot be easily visualised as a whole. One approach is to view the results as a non-linear, interactive document in which only a small part of the results is viewed at any one time. With the addition of simple navigation to select the next sample to view, this approach allows the analyst to easily browse the large result set. More concretely, a one-dimensional sample (a selection of simulations which vary in only one aspect) can be selected from the dataset and visualised as a simple bar chart. Simple rules can then be applied to identify a collection of similar, one-dimensional samples for navigation. To examine this approach, a prototype tool was developed as a web-based application. The basis for this tool was a multi-parameter simulation study of office building energy consumption including 1,440 individual simulations varying across six dimensions including four building types, five building fabrics, three percentages of glazing, the inclusion of daylight control, two glazing types and six HVAC system types (including building load calculations). The tool included a basic report comparing a one-dimensional sample of results and a detailed report showing time series results for an individual case. Navigation panels allowed for simple traversal of the results set and to move between the two reports. The tool was found to be very useful for navigating the multi-dimensional data and the method is generic enough to be transferable to similar datasets.