Abstract
Though distribution system operators have been adding more sensors to their networks, they still often lack an accurate real-time picture of the behavior of distributed energy resources such as demand responsive electric loads and residential solar generation. Such information could improve system reliability, economic efficiency, and environmental impact. Rather than installing additional costly sensing and communication infrastructure to obtain additional real-time information, it may be possible to use existing sensing capabilities and leverage knowledge about the system to reduce the need for new infrastructure. In this paper, we disaggregate a distribution feeder's demand measurements into: 1) the demand of a population of air conditioners, and 2) the demand of the remaining loads connected to the feeder. We use an online learning algorithm, Dynamic Fixed Share (DFS), that uses the real-time distribution feeder measurements as well as models generated from historical building- and device-level data. We develop two implementations of the algorithm and conduct case studies using real demand data from households and commercial buildings to investigate the effectiveness of the algorithm. The case studies demonstrate that DFS can effectively perform online disaggregation and the choice and construction of models included in the algorithm affects its accuracy, which is comparable to that of a set of Kalman filters.
The Challenge of Distribution Grid Visibility
Distribution system operators (DSOs) are increasingly managing "active" grids filled with residential solar and demand-responsive loads, such as smart air conditioners (ACs). To maintain reliability and optimize demand response programs, operators need to know exactly how much power these specific resources are consuming or producing in real-time.
The traditional solution is to increase sensing density, but installing and maintaining communication links for thousands of individual smart meters is prohibitively expensive and often suffers from significant latency. This research addresses the "blind spot" at the substation by asking: how can we take a single, aggregate power signal from a feeder and accurately "unmix" it into its constituent parts (e.g., AC load vs. other loads) as the data arrives?
Methodology: Dynamic Fixed Share for Online Learning
The authors frame disaggregation as an online learning problem. Instead of relying on a single static forecast, they employ a bank of models (experts) and a feedback loop to refine estimates in real-time.
- Dynamic Fixed Share (DFS): This core algorithm maintains multiple models simultaneously and assigns a "weight" to each based on its recent predictive accuracy. If a model starts performing poorly—for instance, a temperature-based model during an unpredicted weather event—the algorithm automatically shifts weight to models that are tracking the real-time feeder data more closely.
- Dynamic Mirror Descent (DMD): Within the DFS framework, DMD is used to adjust each individual model's prediction based on the error between the total predicted demand and the actual measured demand at the substation.
- Implementation Flavors: The study explores two methods: one that treats models as "black boxes" and adjusts their power outputs directly, and another that updates the internal states of dynamic models (like Linear Time-Varying systems) using the real-time error signal, similar to a Kalman Filter.
Experimental Setup and Data
The study validated the DFS approach using a combination of high-fidelity real-world data and simulated feeder architectures:
- Residential Data: 1-minute interval data from the Pecan Street Dataport, covering approximately 2,500 homes in Texas, allowing for ground-truth validation of AC vs. non-AC loads.
- Commercial Data: 4-second interval demand data from municipal and retail buildings.
- Feeder Model: The GridLAB-D R5-25.00-1 feeder taxonomy was used to scale the simulated distribution network.
- Weather Data: Real temperature data from NOAA and Pecan Street were used to drive thermal-dependent AC models.
Performance Evaluation
DFS achieved a Root Mean Square Error (RMSE) for AC demand of approximately 150–200 kW on a feeder with a peak demand of ~12 MW. Crucially, DFS performed comparably to the "Best Kalman Filter" (BKF). While the BKF can only be identified ex post (after the fact), DFS identifies the optimal model combination on-the-fly.
The importance of the feedback loop was evident: when measurement-based updates were disabled, errors increased by nearly 70%. This confirms that real-time feeder feedback is the critical component for accurate disaggregation.
Core Innovations and Future Directions
Unlike traditional "open-loop" load forecasting, this approach is "closed-loop," using the error in the total feeder measurement to refine estimates of hidden components in real-time. The algorithm's robustness was further proven by its ability to ignore "bad" models (e.g., those not accounting for temperature swings) over time as it learned they were inaccurate.
Future work will focus on:
- Automated Parameter Tuning: Moving away from manual heuristics for learning rates and sharing parameters.
- Scaling Dimensionality: Extending the system to simultaneously disaggregate solar generation, electric vehicles, and multiple appliance types.
- Reactive Power: Incorporating reactive power measurements from modern SCADA systems to provide additional features for disaggregation.
- Active Control: Utilizing these real-time estimates as a feedback signal for direct load control, closing the loop between sensing and demand response.
Summary updated