Abstract: This paper presents a hierarchical control framework for a virtual power plant (VPP) that leverages behind-the-meter resources for grid services while maintaining customer privacy during setpoint disaggregation. Unlike many existing approaches, the VPP model uses a hierarchical control strategy and an iterative approach to determine the optimal setpoint disaggregation without direct load control while maintaining system-level power flow and voltage constraints. The proposed approach is numerically validated on a synthetic distribution feeder in San Francisco, demonstrating the ability of the framework to provide privacy-preserving virtual power plant services.
Introduction
As Distributed Energy Resources (DERs)—such as residential batteries, solar panels, and electric vehicles—proliferate, Virtual Power Plants (VPPs) have emerged as a critical tool for aggregating these small-scale assets to provide bulk grid services. However, traditional VPP models face significant hurdles in scalability, distribution network awareness, and customer privacy. Centralized control of thousands of devices is computationally prohibitive, while aggressive VPP dispatch can inadvertently cause local voltage violations. Furthermore, customers are increasingly wary of sharing granular energy usage data or granting third-party operators direct control over their home devices.
This research addresses these challenges by proposing a hierarchical framework that decouples bulk grid demands from local distribution constraints through a multi-layered coordination strategy.
Methodology
The framework utilizes a three-tiered hierarchical structure to manage coordination efficiently:
- Utility/DERMS Layer: The utility employs a LinDistFlow model to calculate feasible power bounds for every bus in the network. This ensures that VPP operations remain within local voltage and thermal limits.
- VPP Operator Layer: The operator receives market signals and uses an optimization model (implemented in Pyomo and solved via HiGHS) to disaggregate these signals into targets for specific geographic subregions.
- Subregion Agent Layer: This layer serves as the privacy-preserving interface. Agents use an iterative disaggregation algorithm to meet subregion targets. By exchanging limited, non-sensitive information with customers in a "request and response" loop, the agent avoids the need for direct access to individual customer data or device control.
Experimental Validation
The framework was validated using the SMART-DS San Francisco Test Feeder, a synthetic urban model representing northeast downtown San Francisco. The feeder features a 10.5 MW peak load and includes 142 residential customer buses serving 306 individual customers, each equipped with a battery system. To test hierarchical coordination, the feeder was partitioned into three subregions using k-means clustering.
Numerical Results and Discussion
Numerical simulations demonstrate the framework's effectiveness in balancing grid services with local constraints. Key findings include:
- Rapid Convergence: The subregion disaggregation error converged within 5 to 15 iterations, confirming the algorithm's viability for real-time grid operations.
- Voltage Stability: Despite aggressive charging and discharging signals, system voltages remained strictly within the 0.95 to 1.05 per unit (pu) safety range.
- Effective Load Shifting: The model successfully managed "virtual batteries" across subregions, showing clear transitions from initial states to post-charge and post-discharge states while respecting individual physical battery limits.
By separating bulk market needs from local network physics, the system allows VPPs to be "network-aware" without requiring the market operator to manage the complexities of every local power line. The iterative approach ensures that the central operator cannot build detailed profiles of individual customer behavior, preserving privacy.
Conclusion and Future Work
This study provides a robust proof of concept for a scalable, privacy-preserving VPP framework. Future research will focus on expanding the model to include a more diverse set of resources, such as rooftop PV and electric vehicles. Additionally, the team plans to integrate federated learning techniques to further harden data security and explore dynamic clustering methods to improve subregion flexibility under changing grid conditions.
Summary updated