Conference Agenda

In 2022, residential buildings in the US accounted for 22% of total energy consumption. The energy use intensity (EUI) of buildings in this sector shows an increasing trend, exacerbated by the effects of climate change. To identify opportunities for improving building energy performance, it is crucial to understand how energy use in residential buildings varies based on climate, socio-demographic factors, and building design parameters. This study utilizes the Residential Energy Consumption Survey (RECS) dataset from the US National Renewable Energy Laboratory (NREL) with 550,000 validated residential energy models, that help understand the different sources of energy residential buildings use and test the energy savings when building parameters are changed. The goal of this study was to take a deep dive into the data through an analytical approach and develop a regression model that can interpret and predict energy consumption patterns of new and existing residential buildings.

The original dataset consisted of 250+ attributes defined through geographical location in the US, socio-demographic conditions including household income and owner type, and building parameters such as height, year of construction, and total built-up area. The dataset pre-processing methods included manual feature selection, followed by exploratory data analysis and feature manipulation. These steps allowed us to reduce the large dataset and identify 53 attributes that acted as the predictors. Normalized Site Energy was chosen as the single response variable of the preliminary study. Twelve regression models were then developed, and their prediction accuracy were compared. Finally, the best-performing model, using a gradient-boosting algorithm, was tested for identifying the priority of energy efficiency measures in a sample residential building.

This study highlights the complexity in data handling due to the presence of qualitative predictors, high dimensionality and multicollinearity of some predictors. The regression models reflected the criticality of tuning the model hyperparameters because of the various factors on which residential energy use relies. The best-performing model acts as a good tool to consult in early stages on potential strategies for energy efficiency in residential buildings without the need for detailed energy modeling. This research can be adopted by building scientist researchers to make more robust prediction tools based on the extensive energy data already available. Such studies have the potential to assist architects, governmental bodies, and policymakers in evaluating optimal strategies for designing future residential buildings with a focus on energy efficiency. Additionally, it provides valuable insights for defining retrofit parameters tailored to enhance energy performance in existing residential structures. This methodology can be expanded further to incorporate specific parameters such as cost, climate zone, building typology, and a detailed breakdown of energy consumption for various end-uses.

By linking a bespoke energy models to the BMSs optimisations can be made automatically to the control system to improve the efficiency of the building by 20-30%. In the past this has been an intensive and highly skilled job with significant human interaction in order to both find the savings and control the building. This presentation will look at the parts of the process that can be automated in order to provide bespoke modelling and continuous optimisations at scale. This presentation will look at how this bespoke model can live alongside the building and be used again and again speeding up retrofit decisions as well as increasing the accuracy of forecast carbon savings from those designs

The presentation will detail some case studies where this automation is being tested.

Traditionally the architecture and engineering fields have worked in silos. One of the fields where this disconnect is more evident is in the field of building performance simulations, where each professional rely only on the analytics they themselves generate without much of a concern for synergies with other fields. It’s hard to look at the entire picture and see if indeed optimization is being achieved or if we’re just transferring inefficiencies from one metric to another. Additionally, many practitioners especially in the architecture field, can’t devise optimization strategies on their designs given the deluge of data that simulations could generate.

In this presentation we will explore the synthetic part of building performance, going beyond the analytics. We will illustrate with case studies how optimization paths could be taxonomized in order of complexity and how, by borrowing concepts from biology & computational design such as sensitivity and fitness, we could incorporate and generate novel design approaches into the architectural design of the projects using building performance data (energy, daylight, solar, etc). In addition, we will show case studies where design optimization concepts such as form finding, performance parametrics, comparative vs. cumulative parametrics and generative design have been explored.

The integration of Large Language Models (LLMs) into the field of Building Energy Modeling (BEM) represents a significant advancement in the application of AI for engineering purposes. This presentation explores the development and application of a context-aware method using LLMs in BEM, with a focus on the analysis of simulation results. We have utilized API calls to OpenAI's GPT 3.5 and 4, demonstrating the adaptability and effectiveness of LLMs in this specialized area.

LLMs are increasingly employed in the scientific community for diverse tasks such as literature review, hypothesis generation, and research summarization. In the context of BEM, we recognized the analysis of simulation results as the most practical and accessible application for LLMs. Our approach departs from conventional LLM applications that are centered on text-in-text-out scenarios. Instead, we target a text-in-numbers-out model, demanding greater precision and uniformity in outcomes to meet the specific requirements of energy engineers and building energy modelers.

To implement this strategy, we developed two agents in Python that make API calls to GPTs: a data extraction/retrieval agent and a data visualization agent. The first agent is tasked with generating SQL queries for extracting data from simulation results, taking into account user queries, specific result file details, and typical characteristics of building simulation result files, such as those from EnergyPlus. The second agent is dedicated to data visualization, merging the context of the extracted data with user requests to produce Python scripts for effective plotting.

This method capitalizes on the strengths of LLMs to yield accurate and dependable results in BEM. The choice of SQL for data retrieval, as opposed to Python, was influenced by our observations that SQL queries offer more consistent and error-free outcomes compared to the broader solution space and potential errors associated with Python scripts.

In conclusion, our study successfully demonstrates the potential of LLMs in enhancing BEM processes through a two-agent system that efficiently handles data extraction and visualization. This approach guarantees precision and consistency in numerical outputs, highlighting the viability of LLMs in specialized domains.

Future Work: Building upon our current success, we aim to extend this methodology to real-time building data, simulating and interacting with Functional Mock-up Units (FMUs), creating simplified control blocks in Modelica, and writing Python scripts that are used with Python plugin of EnergyPlus. These initiatives are expected to significantly elevate the capabilities of building energy modeling and simulation.

Note: Along with the presentation slides, if time allows, we can also demonstrate the application in action and provide a walk-through of the code, if desired.

This presentation will introduce an application of the prototypical building energy models for large scale carbon intensity reduction prediction. This study develops a method to comprehensively assess the long-term carbon intensity reduction potential of aggregated commercial buildings in the continental U.S. This methodology will be demonstrated by the example of K-12 school buildings in the United States.

This presentation will introduce a methodology to systematically create prototypical building energy models for existing buildings. This methodology will be demonstrated by two case studies of creating prototypical energy models for U.S. religious worship buildings and U.S. medium office buildings, representing buildings in 15 climate zones and 2 vintages (pre- and post-1980).

Buildings account for approximately 40% of global energy consumption and associated greenhouse gas emissions, highlighting their critical role in grid decarbonization. The integration of fluctuating renewable energy sources adds significant uncertainties, requiring adaptive energy use in buildings for grid resilience. This adaptation entails a shift from passive consumption to active participation in the energy grid, ensuring demand flexibility and occupant comfort without compromising energy efficiency.

Reinforcement Learning (RL) has emerged as a prominent approach to meet these challenges in building energy management. This presentation examines ten key questions pertaining to the application of RL in flexible energy management of buildings. It covers the impacts of increasing data availability, advancements in machine learning algorithms, the importance of open-source tools, and the practical aspects of software and hardware integration for effective RL implementation.

The presentation aims to provide a succinct introduction to RL in the context of building energy management, offering an overview of current research, pinpointing challenges, and highlighting opportunities for future research directions.