Towards Machine Learning Application in Early-Stage Building Energy Optimization

Licentiate thesis, 2024

As more than one-third of global greenhouse gas emissions are related to the operation of buildings, reducing building energy demand is a key area in the architecture, engineering, and construction (AEC) industry. One promising means is to conduct early-stage building energy optimization. Optimizing the influential architectural design variables (ADVs) in the early stage can significantly reduce building energy demand efficiently. However, current optimization tools often use simulation engines to calculate energy demand for various design alternatives, which can be very time-consuming. To improve the speed of the optimization tools, machine learning (ML) energy prediction models are introduced to replace the energy simulation engines. An ML prediction model requires three crucial elements: the input ADVs, the dataset containing both ADVs and the energy demand of various building design configurations, and the ML algorithm. This thesis investigates how to identify the above elements and provides recommendations for researchers and tool developers at the end.

The input ADVs are identified through a combination of a literature review and a stakeholder survey. Results indicate that building plan, window-to-wall ratio (WWR) on four facades, and wall material are considered influential by both the literature and stakeholders. Building plan, building orientation, shading device, storey number, storey height, roof type, and roof material are also considered influential by stakeholders. It is suggested that insights from both perspectives should be included when developing the ML prediction model. The best-performing ML algorithm, as well as the characteristics of the corresponding synthetic dataset, are identified through multiple ML experiments. A parametric model is developed to generate multiple synthetic datasets with different sizes and different building types, referred to as diversity. Five ML algorithms are selected through a literature review. By conducting ML experiments with different combinations of datasets and ML algorithms, recommendations are provided regarding which algorithms to employ with what type of dataset. Support Vector Model (SVM) performs best in general. Multiple Linear Regression (MLR) performs well with small and low-diverse datasets while Artificial Neural Network (ANN) performs well with large and high-diverse datasets. When generating synthetic training datasets, it is recommended that the dataset needs to have more than 1440 data points and a diversity that covers around 67% of the diversity in the testing dataset to achieve reasonable accuracy.

This thesis explored how ML can support early-stage building energy optimization by investigating how to develop an ML energy prediction model and providing recommendations. This study offers a more holistic point of view in developing ML-based tools by incorporating stakeholders’ opinions in selecting input ADVs. The findings fill the gap of the lack of understanding in the compatibility between different ML algorithms and synthetic datasets in the development of ML models for building energy prediction. In general, this study offers insights for researchers and tool developers in developing ML building energy prediction models and contributes to the application of artificial intelligence in the AEC industry and a more sustainable built environment.