|
"The objective of this project is to create a library of trained diffusion models that generate conditional time series for use in energy system models. Examples include “a 24-hour injection pattern of an offshore wind farm at the NJ coast in spring” or “one week of power consumption of a multi-family residential building in summer.”
Energy system models are essential decision-support tools in engineering, finance, and policymaking. Emerging weather- and behavior-driven energy resources (wind, solar, electric vehicles, heat pumps, etc.) exhibit distinct generation and consumption patterns that must be represented through time series data. Neglecting the time-coupled nature of these resources invalidates most modeling results related to energy storage and resource flexibility, which are two central aspects of modern energy systems. However, access to such data is often limited or requires laborious and sometimes costly data acquisition and preprocessing. Synthetic data generation can overcome this obstacle, and modern diffusion models offer a pathway for generating high-resolution, conditional time series data with manageable computational and training data requirements.
The project has two central tasks: (1) Model training and calibration, and (2) Model dissemination.
Task 1: Each student will receive a time series dataset and train a diffusion model that can synthesize similar time series. This task involves minor data preparation (datasets will be provided by the PI), implementation of the diffusion model (supported by a graduate student in the lab), and calibration/hyperparameter tuning. The latter will be performed in close collaboration with the PI and a graduate student using relevant metrics to evaluate the quality of the synthetic time series (e.g., modes, autocorrelation).
Task 2: Students will publish their model architecture and trained parameters in an open-source repository. This task includes cleaning and documenting the code and creating a public-facing description of the model’s properties. This task will also be closely supported by the PI and a graduate student.
The project goal is for each student to train, calibrate, and disseminate at least one model during the project period. Once the workflow is established, it is expected to be efficiently replicable across multiple datasets. The target is for each student to produce two to three trained models.
Participating in this project will give the students hands-on experience with state-of-the-art machine learning tools and their application to real-world engineering problems in energy systems. They will gain practical skills in data handling, model development, and open-source dissemination. These skills that are increasingly valuable for careers and graduate studies in engineering, data science, and applied mathematics."
|
Sign in
to view more information about this project.
