A Novel Approach for Optimizing Building Energy Models Using Machine Learning Algorithms

Abstract

The current practice with building energy simulation software tools requires the manual entry of a large list of detailed inputs pertaining to the building characteristics, geographical region, schedule of operation, end users, occupancy, control aspects, and more. While these software tools allow the evaluation of the energy consumption of a building with various combinations of building parameters, with the manual information entry and considering the large number of parameters related to building design and operation, global optimization is extremely challenging. In the present paper, a novel approach is developed for the global optimization of building energy models (BEMs) using Python EnergyPlus. A Python-based script is developed to automate the data entry into the building energy modeling tool (EnergyPlus) and numerous possible designs that cover the desired ranges of multiple variables are simulated. The resulting datasets are then used to establish a surrogate BEM using an artificial neural network (ANN) which is optimized through two different approaches, including Bayesian optimization and a genetic algorithm. To demonstrate the proposed approach, a case study is performed for a building on the campus of the Florida Institute of Technology, located in Melbourne, FL, USA. Eight parameters are selected and 200 variations of them are supplied to EnergyPlus, and the produced results from the simulations are used to train an ANN-based surrogate model. The surrogate model achieved a maximum of 90% R2 through hyperparameter tuning. The two optimization approaches, including the genetic algorithm and the Bayesian method, were applied to the surrogate model, and the optimal designs achieved annual energy consumptions of 11.3 MWh and 12.7 MWh, respectively. It was shown that the approach presented bridges between the physics-based building energy models and the strong optimization tools available in Python, which can allow the achievement of global optimization in a computationally efficient fashion.