Expanding Performance-Driven Parametric Design Spaces Through Data Streams

There has been recent interest in using sampled datasets of parametric models to create fast, accurate predictions of building performance. This rapid feedback can enhance creative processes in which a designer is exploring potential design options live, compared to waiting for the results of individual simulations for each design possibility. However, basing these predictions on an existing parametric dataset can be limiting, as the original variable structure must be maintained. Drawing from advances in data streams, this paper proposes and evaluates strategies for adding a variable to the parametric design space without needing to re-simulate the entire dataset to update the prediction. Several approaches are tested to expand the design variables for three situations: a daylit room, a massing model for energy prediction, and a stadium geometry. The main strategy, an online updating method, significantly reduces the number of new simulations required with the new variable to achieve reasonably accurate performance prediction, compared to re-simulating the entire dataset with the new variable included. However, in certain cases this approach only slightly outperforms creating a simple linear model with just the new datapoints. Future work can continue to investigate how to best enable the addition of variables to a parametric design space and corresponding performance prediction model.