Improvements to building energy usage modeling during early design stages and retrofits

Abstract

A variety of improvements to the MIT Design Advisor, a whole-building energy usage modeling tool intended for use during early design stages, are investigated. These include changes to the thermal mass temperature distribution and lighting requirement calculations, the formulation of a representative day scheme to reduce the number of days that need to be simulated to find energy usage for a full year, the creation of an optimization algorithm to allow users to improve on their designs, and the addition of an algorithm to predict potential savings from retrocommissioning (RCx) using a limited set of simple inputs. Design Advisor itself is also tested for usability, speed, and accuracy using three existing buildings. The frequency of thermal mass-related calculations is reduced by finding the limits of the semi-implicit Crank-Nicolson method before it begins to return physically implausible oscillatory temperature profiles. An effort is made to speed up lighting calculations using a multivariate regression analysis in place of a multiple-reflection-based illuminance model. Representative days are formed by creating an average and two extreme weather days per month using existing climatological data, reducing the number of simulated days per year from 365 to 72 (three per month, repeated once for training). Combined, these changes lead to reductions in run time of up to 50% with roughly 10% loss of accuracy. The optimizer leverages these run time improvements to rapidly find optimal building designs given a set of input constraints. Initially, a multistep multivariate regression is used to reduce the given search space and tighten the constraints. Then, a genetic algorithm is used to find the target solution. Initial tests of this combination have led to average reductions in energy usage of 25% given 6 minutes of calculation. To extend Design Advisor's applicability to existing buildings, an algorithm for predicting potential energy savings from RCx is implemented and tested. A database of 90 buildings that have undergone an RCx process and had their resulting energy savings documented has been collected. A k-nearest neighbors algorithm is used to evaluate the potential savings of test buildings based on this data set, operating on the assumption that similar buildings (in terms of location, size, and energy usage intensity) will present similar faults or opportunities for savings. While the average savings percentage prediction error is 0.02, the root-mean-square error is 12.4, which is greater than the actual savings potential of many buildings. Model validation is performed using three existing buildings; two in the Philadelphia area and one on MIT's campus. For energy types for which no building faults or other issues were later found (as in the MIT building), final usage predictions are found to be accurate to within a mean bias error of 11.2% to 2.6%. To improve upon these accuracies, further details about key building parameters and modes of operation would be required. These studies also inform further usability improvements, including reporting site (rather than primary) energy usage and expending reported electricity usage to include loads other than lighting.