Open amahajan68 opened 1 day ago
9/16
to 9/22
Description: Investigated inputs required for the multivariate regression model, specifically focusing on compactness and other key features, to determine if available data can be used to recreate these inputs for energy prediction.
Link to Work:
Outcome & Reflection: We identified that inputs such as compactness can potentially be recreated using existing data, which simplifies the process of input preparation for the regression model. This realization also opens the door for using synthetic datasets based on building geometry and footprint.
Meeting Highlights:
Discussed project titles:
Identified potential users of the surrogate model:
Architects and government planners who require quick simulations to assess the energy impact of design changes at scale.
Discussed the problem definition:
The challenge of modeling energy performance for a large number of buildings at an urban scale.
Emphasized the need to combine white-box simulations (detailed, physics-based) with black-box models (machine learning) to create a scalable, grey-box solution.
Explored neighborhood-scale tools such as UMI for integrating energy performance across multiple buildings.
TODOs from the meeting:
Explore UMI and request Georgia Tech building data (Joseph).
Build a 3D model for a specific Atlanta neighborhood.
Use OpenStreetMap for building footprint and height data.
Investigate the Google Dataset for energy performance (Anubha).
Look into tools that can extract window-to-wall ratio or other building geometry details.
Peer Feedback:
9/23
to 9/29
Customized Simulation Approach: The paper presents a lightweight simulator customized for each building by calibrating it to telemetry data. The simulator achieved only 0.64°C drift over six hours, making it highly accurate for real-world use.
RL for HVAC Control: The RL agent optimizes HVAC energy consumption by adjusting setpoints to balance energy use, carbon emissions, and occupant comfort.
Simulator Configuration & Scalability: The simulator is easily configurable for different buildings using floorplans and HVAC metadata. While currently designed for single buildings, the paper highlights potential scalability to city-scale energy management.
Calibration Method: Finite Differences (FD) is used for modeling thermal diffusion, ensuring accurate simulation of heat transfer across the building. Real-world telemetry data is used to calibrate the simulator.
Performance: The simulator demonstrated Mean Absolute Error (MAE) as low as 0.64°C, validating its accuracy across different time intervals and conditions.
Future Work: The authors plan to improve generalization by tuning across multiple seasons and expanding the action space to include additional HVAC parameters.
Potential for Expansion: The approach can be adapted for multi-building environments, offering the potential for city-wide energy optimization.
Description: I reviewed Google's TensorFlow Smart Buildings Simulator and the accompanying SMART Buildings Dataset, which includes six years of telemetry data from three commercial office buildings. I also examined the paper from BuildSys '23, "A Lightweight Calibrated Simulation Enabling Efficient Offline Learning for Optimal Control of Real Buildings."
Link to Work: Google Paper on Smart Buildings
Outcome & Reflection: The paper provided key insights into RL-based HVAC optimization and calibration techniques. The simulator achieves less than 0.64°C drift over six hours, making it highly effective for single-building energy management. There is potential for adapting the model to simulate multiple buildings for larger-scale energy optimization solutions.
Description: Set up a Linux environment using Oracle VirtualBox to run the Google TensorFlow Smart Buildings Simulator. This provided the necessary dependencies for running the model in a Linux-based environment.
Outcome & Reflection: After troubleshooting performance issues related to the virtual machine, switching to Oracle VirtualBox resolved the issues. The Linux environment is now fully operational for running the simulator and making further modifications.
Current Status: Currently researching the RL models used in the Google simulator and focusing on how they can be scaled to optimize energy consumption across multiple buildings.
Expected Completion: Next week
Takeaway: Combining physics-based simulations with real-world data calibration is highly effective for RL training in HVAC systems. Expanding the simulator for multi-building optimization could unlock new possibilities in urban energy management.
Skills Acquired:
9/30
to 10/6
Description: Explored the process of transforming Comstock IDF files to extract geometric features such as window-to-wall ratio, area, and other building attributes. The goal is to automate this workflow and integrate it with our multivariate regression model for energy consumption analysis.
Current Status: The initial steps for extracting geometric data from Comstock IDF files were completed. We are now working on automating the workflow for large datasets.
Expected Completion: Monday
Weekly Notebook Entry — Week 4
Overview
9/9
to9/15
Tasks for This Week
Tasks
Task: Investigate Commercial Building Features - (completed)
Task: Overview of Previously Used Models - (ongoing)
Task: Create a Pipeline for Multivariate Regression Model - (completed)
Description: I used the Mockaroo website to generate mock data, which was then input into the Multivariate Regression Model. This mock data simulated different building parameters like glazing area, surface area, wall area, relative compactness, and more. The purpose was to ensure we have a fully functioning pipeline for energy consumption predictions based on these parameters.
Link to Work:
Outcome & Reflection:
Challenges & Solutions
Challenge: Determining whether we need a new model and understanding the inputs for the multivariate regression model. The question of whether we can use ComStock data for the regression model remains.
Challenge: Generating realistic input data for the multivariate regression model.
General Takeaway, Skills & Knowledge Acquired
Takeaway: IDF files from DOE reference buildings might allow us to reconstruct missing data for the regression model and help simulate buildings for Atlanta-specific energy predictions. Additionally, using Mockaroo as a tool for generating mock data ensures flexibility and scalability in testing the model.
Skills Acquired:
Team Meetings
Meeting Highlights:
Peer Feedback: Focus more on automating parametric models and integrating them into the regression pipeline.
Personal Reflection
Additional Notes & Resources
Plans for Next Week