Search

Yun-Dam Ko / 고윤담

Building (Data) Scientist

“Make buildings and cities more sustainable by bridging data and building systems through scientific understanding”

yundam.ko@gmail.com

Academic CV

CV-YUN-DAM_KO.pdf
305.6 KiB

Research Interests

#Building Informatics #Building Performance Simulation #(Urban) Building Energy Modeling

NEWS

JUN 22-26, 2026 NOV 19-21, 2025 SEP 15, 2025
Heading to Banff! Thrilled to present our research at BuildSys 2026 soon. Attended BuildSys 2025 in Golden, CO, USA Thrilled to be selected as a Stanford HAI Graduate Fellow!
NEWS Archive

PROFESSIONAL EXPERIENCE

2026 -
2024 -
Student Assistant
•
Building Technology Urban Systems Division
Graduate Research Assistant
•
AI for Sustainable Building Systems
Graduate Teaching Assistant
•
CEE 156/256: Building Systems Design & Analysis
2023 - 2024
NINEWATT, South Korea
Lead Research Engineer (Mar 2023 - Sep 2024)
•
Developed ChatGPT-based time-series forecasting model and building systems control strategies (Managing collaborative project with OpenAI)
Snapshot
•
Developed automated (urban) building energy modeling tool
Demo
•
Developed building-mobility energy simulator (Managing pilot project with Hyundai Motor Group)
Demo
•
Co-developed of urban building energy data analysis platform, WATTI
Snapshot
2021 - 2024
Military Duty
•
Alternative military duty service as technical research personnel
2021 - 2023
TENELEVEN Inc., South Korea
Software Developer
•
Developed outdoor airflow simulation solver using OpenFOAM and deep learning (GAN)
Snapshot (gif)
•
Co-developed PV energy analysis tool using RADIANCE and machine learning
Snapshot #1: Solar radiation analysis
Snapshot #2 (gif): PV analysis
•
Managed and coordinated software localization project (Singapore)
Snapshot
2019 - 2021
Research Assistant
•
Research Projects:
◦
Data-Centric Checkup Technique of Building Energy Performance
◦
Development of Smart Hybrid Envelope Systems for Zero Energy Buildings through Holistic Performance Test and Evaluation Methods and Fields Verifications
◦
Development of Real-time Diagnosis Technology of Home Energy Usage and Smart & Autonomous Control/Management System
2018 - 2019
Research Assistant / Thesis
•
Developed generic input parameters for DX cooling coil simulation model in whole building simulation software, WUFI Plus
•
Created whole building simulation models for passive houses with WUFI Passive
•
Assisted in evaluating practicability of urban climate simulation model

EDUCATION

SEP 2024 -
AUG 2021
Ph.D. in Civil Engineering
Thesis: Parameter estimation of unknown building properties using transfer learning (Advisor: Prof. Cheol-Soo Park)
SEP 2018
B.S. in Civil Engineering
Thesis: Generating generic input parameters for a DX cooling coil simulation model (Advisor: Prof. Klaus Peter Sedlbauer)

PROJECTS

•
Snapshot
•
The developed model was deployed and served using AWS Lambda in BUILDIT DESIGNER, a semi-automated 3D modeling computer-aided design program for architecture applications.
•
Abstract
Although CFD (Computational Fluid Dynamics) simulation has been widely adopted to investigate urban wind environments, its high modeling difficulty and high computational cost hinder its integration into urban and architectural design process, especially in the early design stage. This study presents a generative model to synthesize CFD simulations for the real-time prediction of the wind velocity field around buildings. A generative adversarial network, Pix2Pix, is trained on the CFD simulation results of 1,146 building configurations. As a result, the developed model can accurately generate an wind velocity field in an urban area up to 700,000 times faster than the conventional CFD model.
•
Development process of the generative model for urban wind environment simulation
Generative model for urban wind environment simulation
•
◦
To achieve an ambitious target in carbon emission reduction, it is essential to explore ways to save energy in the building sector through energy retrofitting and optimizing operations, among other strategies. However, conducting retrofit analysis and operational optimization for even a single building demands expert audits and in-depth analysis, entailing substantial effort and cost. This fact leaves numerous small to medium buildings, which occupy a significant portion and possess high energy-saving potential, neglected without reaping the technological benefits. Therefore, the aim of this project is to develop an automated framework for building energy modeling in an urban context, enabling users to easily and rapidly create simplified whole-building energy models, not only for individual buildings but also for multiple buildings simultaneously. The anticipated outcomes will find applications in carbon reduction policy-making, solar analysis (e.g., PV generation), and retrofit analysis. The target users include local governments, owners of small to medium-sized buildings, architectural firms, and construction companies, who seek cost-effective solutions without incurring significant expertise expenses.
Automated Urban Building Energy Modeling using GIS data
•
Abstract
This study proposes a transfer learning (TL)-based inverse modeling to identify unknown building properties. This study examines the transfer from virtual buildings to existing buildings, especially for identifying wall U-value, HVAC efficiency, and lighting power density. For this purpose, synthetic data was generated from simulation results of sampled EnergyPlus models, and then we developed artificial neural network (ANN) models using this data. By adopting TL, the ANN models were transferred to the domain of existing buildings, and the performance of TL was tested on 61 existing buildings. It was found that the transferred ANN models can rapidly identify unknown factors in existing buildings. It is noteworthy that data collected from EnergyPlus simulation runs can improve the performance of a data-driven model through TL. It is expected that the use of TL enables the developed model to be reusable for another group of buildings with improved performance and reduced training time.
•
Workflow
•
Baseline vs. proposed method (TL)
•
Findings
◦
the data-driven model trained only with virtual buildings’ data cannot be directly used for existing buildings, as we expected. However, the models trained with data of both virtual and existing buildings through TL significantly outperform the models trained only with data of existing buildings.
◦
The TL procedure can be conducted in a computationally efficient manner (about 16 seconds in this study), because only a few parameters (17 parameters in this study) are re-trained, i.e. easily scalable to a large group of buildings.
◦
Data collected from the simulation results of a physics-based model (e.g., EnergyPlus) can be utilized to improve the performance of a data-driven model by TL.
◦
The performance achieved in this study can be further improved by re-transferring the developed model to other buildings. In other words, the developed TL models in this study can be reused as a pre-trained model and transferred to another group of buildings, while maintaining useful knowledge captured from buildings used in this study (See Figure below).
Expected performance improvement of transfer learning according to number of transfers
Transfer learning to identify unknow building properties from virtual to existing buildings
•
Introduction
◦
The aim of this project is to propose a methodology for probabilistically exploring scenarios that achieve a energy-saving target in buildings.
◦
As a case study, the proposed methodology was applied to investigate how three occupant behaviors (Cooing set point temperature [℃], Cooling start temperature [℃], and CO2 concentration [PPM] when opening/closing windows) should be modified based on two different scenarios to achieve the target of 10% or 20% energy savings in a residential building located in Seoul, South Korea.
▪
Scenario #1: Heavy-consuming households save more than light-consuming households proportionally to achieve 10% or 20% of the mean energy saving.
▪
Scenario #2: All households in the residential building equally reduce energy consumption by 10% or 20%.
•
Method
◦
Target building: 87 households in a residential building located in Seoul, South Korea
(a) View from North (b) View from West (c) Floor plan (55 m2) of households
◦
Building energy modeling using EnergyPlus
▪
Simulation inputs
◦
Occupant behavior modeling (rule-based)
▪
It is assumed that each agent (i.e., occupant) in a household follows the rules defined in this study
◦
Occupancy schedule collected by the field survey
Occupancy schedule of a household out of 87 households
◦
Surrogate modeling using Artificial Neural Networks (ANN) to reduce computation time
•
Result
◦
The likelihood distribution: baseline (black) vs. scenario #1 (red) vs. scenario #2 (blue)
◦
The posterior distribution of each occupant behavior
▪
10% saving target
•
The result shows that the 10% savings target could be achieved by adjusting cooling set point temperatures alone to a higher level, as indicated by the posterior distributions.
•
The posterior distributions of the other two occupant behaviors (cooling start point temperature and window open/close) are multimodal. This might be because these behaviors are relatively insensitive and/or correlated to the other involved behaviors, making it difficult to identify an adjusting direction.
▪
20% saving target
•
The result shows that the 20% savings target could be achieved by adjusting cooling set point temperatures for Scenario #1 and adjusting all three occupant behaviors for Scenario #2, as indicated by the posterior distributions.
◦
The posterior predictive distributions
▪
Based on the posterior predictive distributions, the success probability of each scenario and each target can be quantified.
▪
10% saving target
▪
20% saving target
◦
Conclusion
▪
Bayesian inference can support decision-making regarding energy policy in a city with appropriate risk management.
▪
The posterior of three occupant behavior measures were estimated, given the observations of the two different scenarios of two energy saving targets (10% and 20%).
▪
If an estimation parameter is correlated to the others, it could be difficult to make a decision (in this study, whether open window more often or not). We should be careful when selecting estimation parameters to obtain an accurate result.
▪
In this study, the priors are assumed to be uniformly distributed (my belief), and this affected the posteriors. Define an actual prior (e.g. domain knowledge, or data) may be important to improve the reliability of the posteriors.
Bayesian inference to identify occupants’ energy saving behavior in a residential building
•
Abstract
Because occupant behavior (OB) is one of major sources of the performance gap between simulated and actual building energy use, many studies have focused on developing data-driven OB models. However, developing accurate OB models demands large amount of detailed data containing occupant information. Unfortunately, collecting such occupant data is difficult due to privacy and sensor installation. This study aims to present an inverse approach to estimate OB from measured end-use energy data, or hourly aggregated electricity consumption data of a household. For this purpose, occupant behavior patterns (cooling set-point, cooling start temperature, window open/close) were propagated (27 classes, 42,000 samples) and EnergyPlus, one of the most sophisticated dynamic simulation tools were employed. The developed residual neural network (ResNet) model is successful (accuracy level = 86.3%) for identifying the aforementioned occupant behavior patterns.
•
Workflow
•
Data praparation
◦
Target building: 87 households in a residential building located in Seoul, South Korea
(a) View from North (b) View from West (c) Floor plan (55 m2) of households
◦
Building energy modeling using EnergyPlus
▪
Simulation inputs
◦
Occupant behavior modeling (rule-based)
▪
It is assumed that each agent (i.e. occupant) in a household follows the rules defined in this study
◦
Occupancy schedule collected by the field survey
Occupancy schedule of a household out of 87 households
•
Training and Validation
◦
Validation accuracy of ResNet model for identification of three occupant behaviors
◦
Accuracy evolution of ResNet model for 27 occupant behavior patterns
ResNet to identify occupant behavior patterns from aggregated household energy data
•
Summary
In the forward approach where design parameters are input and EUI is output, the number of possible design solutions is too many to conduct an exhaustive search because multiple continuous and discrete design parameters are involved. In addition, the impact of each design parameter on EUI must be identified for time-efficient search of design solutions. In contrast to the forward approach, target EUI is input and feasible ranges of design parameters are output in the inverse approach. The inverse approach can be used for rapid exploration of design solutions and rational decision making. For this purpose, this case study was intended to generate feasible ranges of design parameters that meet a target EUI (90 kwh/m2/yr) and a 30% energy savings target of an existing education building. As a result, what follows is found: (1) sensitivity ranking of significant design parameters of the existing building, (2) feasible ranges of design parameters meeting target EUI, and (3) inferred energy performance of the existing building after energy retrofit.
•
Motivation
•
Methodology
•
For more detailed, visit the followings:
Inverse approach for building energy retrofit analysis
•
Snapshot
◦
Real-time masking others’ face except for my face
◦
Application case: YouTube
•
Abstract
With the growing popularity of online video and streaming platforms, such as YouTube and Twitch, watching videos or live streams has become a common part of our daily life. However, without the proper privacy protection, private lives of both the public and online streamers can be revealed through shared videos and live streaming. It occurs not only due to the lack of privacy knowledge, but also due to the lack of privacy protection tools that allow users to filter sensitive information without much burden.
In this regard, the aim of this project is to present a deep learning-based privacy protection tool which can be used to filter pedestrians’ faces in videos or live streams in real time. Two major issues should be considered when building the tool: (1) streaming or sharing videos should not be transmitted to the cloud for inference to prevent possible privacy abuses, and (2) computation cost of the tool should be low enough to be in real time realm. To overcome these issues, the authors leveraged light weight deep neural networks, such as MobileNetV2 and SSD (Single shot detector), and Google Coral Dev Board, a commercially available development board containing Google Edge TPU coprocessor – a 8-bit fixed point hardware.
•
Development process
•
Model
◦
SSD MobileNet V2 for face detection
▪
Tensorflow object detection API + Face dataset (WIDERFACE)
◦
The model was quantized using Tensorflow lite to deploy on Google Edge TPU (Coral dev board)
Real-time face masking on Edge: Privacy preservation in videos
•
Abstract
Energy benchmarking has been used to compare energy efficiency of buildings. The conventional benchmarking uses Energy Use Intensity (EUI, kwh/㎡·yr) per ‘use type’ (e.g., residential, office, and educational). However, it could lead to a false comparison when two buildings of a same ‘use type’ have different thermal characteristics such as HVAC, building envelopes, and operational hours. We conducted K-means clustering analysis with regard to monthly energy data of 10,750 buildings. It was found that those buildings can be clustered into four representative patterns. Then, we compared the clustering results of all buildings and buildings of each use type. It was found that the representative energy use pattern of each use type was very similar to it of all use type, meaning that a building use type might not reflect the monthly operation pattern. Therefore, it could be more meaningful to compare two buildings of a same pattern than of a same use type.
•
Analysis process
•
Find the optimal K using Silhouette score
◦
When K=2 respectively for both electricity and gas consumption data, each building is optimally matched to its own cluster.
•
Representative four energy use patterns
•
Clustering energy data of each building use type
◦
The representative energy use pattern of each use type was very similar to it of all use types together.
Clustering analysis of monthly energy use of 10,750 buildings for benchmarking
•
Abstract
Although numerous studies have been conducted on performance gap between predicted and measured energy uses, there is a significant lack of evidence regarding whether monthly energy use of each building repeats over time. This study aims to evaluate the repeatability of monthly energy use of 10,750 buildings for three years (2017-2019). It is assumed that the higher CVRMSE of building's monthly energy use between a pair of years is, the less repeatable a building is, leading to being less predictable. K-means algorithm was used to extract representative energy use patterns of the buildings, and four energy use patterns were identified. It is noteworthy that 9,760 (90.8%) buildings do not satisfy the monthly criteria of ASHRAE Guideline 14 (CVRMSE=0.15), which means their three years' monthly energy use pattern do not repeat. This signifies that it might be difficult to predict the energy use of those buildings. In addition, it was found that there is a distinct difference of CVRMSE (i.e., the degree of unpredictability) among four energy use patterns (0.27, 0.28 vs. 0.38, 0.40 on average), whereas there is little difference among seven building use types (0.27-0.31 on average).
•
Metric to quantify repeatability of monthly building energy use
•
How repeatable is dry-bulb outdoor temperature of Seoul, South Korea?
◦
The average CVRMSE of all pairs of three years is 0.13, meaning that the outdoor temperature of these three years are relatively repeatable.
•
How repeatable are monthly energy use of 10,750 buildings?
◦
Only 990 (9.2%) out of 10,750 buildings meet the monthly calibration criteria of ASHRAE Guideline 14 (CVRMSE ≤ 0.15) for three years.
▪
For the other 9,760 buildings, even if a building energy model is validated on a 12-month basis, its performance cannot be guaranteed for next several years.
◦
Sample building
Repeatability of monthly energy use for 10,750 buildings

PUBLICATIONS

Peer-reviewed Journal Articles
Peer-reviewed Conference Proceedings
Domestic Conference Proceedings (in Korean)

HONORS AND AWARDS

2026-2027
2026
2025-2026
2020.12
2020.11
2020.09 - 2021.08
2020.03 - 2021.02
Emerging Scholars Network Stanford Office of the Vice Provost for Graduate Education
ACM SIGs Student Travel Grant Association for Computing Machinery
Stanford HAI Graduate Fellow The Stanford Institute for Human-Centered Artificial Intelligence (HAI)
Poster Session Award Winner (Runners-Up) The 2020 Winter Simulation Conference
Outstanding Paper Award The 2020 Autumn Annual Conference of the Korean Institute of Architectural Sustainable Environment and Building Systems
BK (Brain Korea) 21 Four Scholarship The National Research Foundation of Korea
TAEJAE Research Scholarship TAEJAE research foundation

PATENTS

1.
Kim, Y.R., Park, S., and Ko, Y.D. (2024). Method of forecasting the power generation of a power generation facility. Patent No. 10-2799556 (South Korea)
2.
Kim, Y.R., Park, S., and Ko, Y.D. (2023). Automation method and system for building energy modeling. Patent No. 10-2708831 (South Korea)
3.
Kim, Y.R., Park, S., Ko, Y.D., and Kim, R. (2023). Server for selecting buildings in need of energy efficiency improvements and method thereof. Patent No. 10-2708834 (South Korea)
4.
Kim, Y.R., Park, S., Ko, Y.D., and Kim, R. (2023). Method for Assessing Energy Use Information at a Specific Address and Its Associated Server. Patent No. 10-2799577 (South Korea)
Made with