A Machine Learning-Based Approach to Predict and Optimize the Performance of Zero Energy Building (ZEB): A Case Study for Florida

Abstract

Machine learning is currently one of the most searched fields aiming to solve real-life problems. Building simulation software tools help engineers estimate building energy behaviors before the actual construction, allowing implementation of more energy efficient choices in building design and construction. Current building energy simulation software tools are mostly physics-based and still lack the benefit obtained with machine learningbased modeling, which offers fast and less computationally expensive techniques to build energy models and efficiently perform design optimization. This thesis presents a machine learning-based approach for building energy modeling and optimizing design parameters to minimize building’s energy consumption. The study is comprised of three main stages. These include creating an EnergyPlus simulation model to generate a physics-based model for the building with all building characteristics. The model is used to generate a database containing input design parameters that are used in energy modeling and annual energy consumptions for different energy models. The results obtained from this database are then used in the second stage, which involves developing an artificial neural network-based surrogate model. The neural network performs simulations by taking a set of inputs and trying to predict an output. The inputs, in this case, are building design parameters and control settings, while the output is the building energy consumption, photovoltaic system power production, and the corresponding net site energy. The third stage is the optimization stage implemented on the surrogate model to determine optimal design variables that provide minimal energy consumption. Design parameter search space along with the surrogate model are provided as inputs to the optimization algorithm. The study uses two different optimization approaches, including the genetic algorithm and the Bayesian method. This iv study shows that the proposed machine learning-based strategy accurately estimates overall energy usage and production. Furthermore, the model optimization is implemented on the neural network at far less computational costs and time than the traditional strategies that involve numerous co-simulation tools to obtain the same results. The developed approach bridges between physics-based building energy models and strong optimization tools available in python which can allow achieving global optimization.