Implementation of Learning-Based Dynamic Demand Response on a Campus Micro-Grid

Sanmukh R. Kuppannagari, Rajgopal Kannan, Charalampos Chelmis, Viktor K. PrasannaIJCAI 2016 · OpenReview

Abstract

Demand Response (DR) allows utilities to curtail electricity consumption during peak demand periods. Real time automated DR can offer utilities a scalable solution for fine grained control of curtailment over small intervals for the duration of the entire DR event. In this work, we demonstrate a system for a real time automated Dynamic DR (D2R). Our system has already been integrated with the electrical infrastructure of the University of Southern California, which offers a unique environment to study the impact of automated DR in a complex social and cultural environment including 170 buildings in a "city-within-a-city" scenario. Our large scale information processing system coupled with accurate forecasting models for sparse data and fast polynomial time optimization algorithms for curtailment maximization provide the ability to adapt and respond to changing curtailment requirements in near real-time. Our D2R algorithms automatically and dynamically select customers for load curtailment to guarantee the achievement of a curtailment target over a given DR interval.

Introduction

The stability of a power grid relies on a delicate, constant balance between supply and demand. Demand Response (DR) is a strategy where utilities request customers to reduce electricity usage during peak periods to prevent grid overload. However, traditional DR programs are often "static," relying on schedules communicated far in advance. This leads to inefficiencies because predictions made hours or days ahead are often inaccurate.

Researchers at USC have developed and deployed a real-time, automated "Dynamic Demand Response" ($D^2R$) system on a campus-wide micro-grid. Managing a complex environment like a university campus—a "city-within-a-city"—involves hundreds of buildings with varying usage patterns and often sparse sensor data. This research addresses the challenge of making DR "dynamic," reacting in near real-time (15-minute intervals) to ensure curtailment targets are met without causing significant discomfort to occupants.

System Architecture

The core of the $D^2R$ system is the Policy Engine (PE), which manages the flow of information between the utility, campus facilities, and individual building meters.

System architecture showing the interaction between the Utility, Facilities Management, and the Policy Engine.

The Policy Engine consists of two primary components:

  1. Prediction Engine: This module utilizes historical data and time-series forecasting (ARIMA and Lasso-Granger) to learn occupant energy preferences. A key innovation is the "influence-based learning model," designed to handle missing or intermittent data. It can accurately predict campus-wide consumption using data from only approximately 7% of the total meters.
  2. Optimization Engine: Once predictions are established, the system determines which buildings should curtail power and which strategies to employ (e.g., resetting temperature set points or reducing airflow). This is framed as a Sustainable DR (SDR) problem, distributing the curtailment target proportionally across the DR interval to avoid aggressive, unsustainable cuts.
The internal logic loop of the Policy Engine, connecting prediction and optimization with observed data.

The system integrates with campus infrastructure using the OpenADR standard to receive curtailment targets from the utility and provide acknowledgments.

Data and Experimental Setup

The study utilizes real-world data from the University of Southern California (USC) campus micro-grid, encompassing 170 buildings instrumented with smart meters. Data is aggregated into 15-minute intervals. The dataset includes historical consumption, weather data, and schedule data to train the predictive models.

Key Insights and Results

The $D^2R$ system demonstrated several significant advantages over traditional methods:

  • Sparse Data Robustness: By using influence-based models that understand how different buildings' energy patterns relate, the system can predict whole-campus behavior using only a ~7% subset of real-time meter data.
  • Sustainability and Precision: The SDR approach ensures curtailment is distributed fairly. In testing, the algorithms achieved a very low absolute error range of 0.001-0.05 kWh.
  • High Impact and Accuracy: The system successfully achieved curtailment values as high as 1.2 MW in a single DR event. It maintained curtailment errors at $\le 10\%$, a significant improvement over the 30% error rates seen in contemporary techniques.
  • Dynamic Adaptability: The optimization algorithms are fast enough (polynomial-time) to run in near real-time, allowing the system to adapt every 15 minutes if a building becomes unresponsive or hits a thermal comfort limit.

Conclusion

While the $D^2R$ system excels in a controlled campus environment, broader adoption faces challenges. Moving to a city-wide scale introduces the "human factor," requiring effective financial or social incentives for residential and commercial compliance. Furthermore, scaling to tens of thousands of buildings will require even more robust approximation methods, and tighter integration with meteorological data will be necessary to handle environmental volatility.

Summary updated

Abstract

Demand Response (DR) allows utilities to curtail electricity consumption during peak demand periods. Real time automated DR can offer utilities a scalable solution for fine grained control of curtailment over small intervals for the duration of the entire DR event. In this work, we demonstrate a system for a real time automated Dynamic DR (D2R). Our system has already been integrated with the electrical infrastructure of the University of Southern California, which offers a unique environment to study the impact of automated DR in a complex social and cultural environment including 170 buildings in a "city-within-a-city" scenario. Our large scale information processing system coupled with accurate forecasting models for sparse data and fast polynomial time optimization algorithms for curtailment maximization provide the ability to adapt and respond to changing curtailment requirements in near real-time. Our D2R algorithms automatically and dynamically select customers for load curtailment to guarantee the achievement of a curtailment target over a given DR interval.
BibTeX
@inproceedings{kuppannagari2016implementation,
  author = {Sanmukh R. Kuppannagari and Rajgopal Kannan and Charalampos Chelmis and Viktor K. Prasanna},
  title = {{Implementation of Learning-Based Dynamic Demand Response on a Campus Micro-Grid}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
  url = {https://openreview.net/forum?id=BJbuCzMOWB},
}
Figure