Data-Driven Decentralized Optimal Power Flow

Roel Dobbe, Oscar Sondermeijer, David Fridovich-Keil, Daniel Arnold, Duncan S. Callaway, Claire J. Tomlin

Abstract

The implementation of optimal power flow (OPF) methods to perform voltage and power flow regulation in electric networks is generally believed to require extensive communication. We consider distribution systems with multiple controllable Distributed Energy Resources (DERs) and present a data-driven approach to learn control policies for each DER to reconstruct and mimic the solution to a centralized OPF problem from solely locally available information. Collectively, all local controllers closely match the centralized OPF solution, providing near optimal performance and satisfaction of system constraints. A rate distortion framework enables the analysis of how well the resulting fully decentralized control policies are able to reconstruct the OPF solution. The methodology provides a natural extension to decide what nodes a DER should communicate with to improve the reconstruction of its individual policy. The method is applied on both single- and three-phase test feeder networks using data from real loads and distributed generators, focusing on DERs that do not exhibit inter-temporal dependencies. It provides a framework for Distribution System Operators to efficiently plan and operate the contributions of DERs to achieve Distributed Energy Services in distribution networks.

The Challenge of Grid Modernization

As the global energy landscape shifts toward sustainability, distribution grids are experiencing a massive influx of Distributed Energy Resources (DERs), such as residential solar panels and electric vehicles. While these resources offer significant potential for grid support, managing their impact on voltage and power flow is a complex task for utilities.

Traditionally, this management is handled through Optimal Power Flow (OPF), a centralized optimization problem that requires a comprehensive, real-time view of the entire network. However, most distribution grids lack the high-speed communication infrastructure necessary to transmit data from every node to a central controller and back. Existing decentralized alternatives, such as simple volt-var curves, often lack the precision needed for optimal performance and are difficult to tune. This research asks a fundamental question: Can we train local controllers to "guess" the optimal centralized solution using only the data available at their specific location?

Methodology: Learning to Mimic Centralization

The authors propose a four-step framework designed to decentralize OPF through machine learning, effectively allowing local resources to mimic a global optimizer.

  1. Data Collection: The process begins by gathering historical data on loads and generation, often sourced from smart meters and weather sensors.
  2. Centralized Benchmarking: A full, centralized OPF is run on this historical data to determine the "perfect" setpoints ($u^*$) for every DER in the system.
  3. Information-Theoretic Analysis: Using a Rate Distortion Framework, the researchers treat decentralization as a data compression problem. This allows them to quantify how much global grid information is implicitly contained within local measurements.
  4. Supervised Learning: Each DER is equipped with a multiple stepwise linear regression model. These models map local exogenous variables—factors the controller cannot change, like local load and PV generation—to the optimal setpoints identified in the benchmarking phase.
The proposed multi-agent system architecture, showing the flow from data gathering to centralized optimization and finally to decentralized learning.

By utilizing nonlinear features, such as quadratic and interaction terms, these models capture the complex physics of the grid while remaining computationally efficient enough to run on low-cost embedded hardware.

Key Insights: The Compression Analogy

A core innovation of this work is the comparison of grid decentralization to image compression (like the JPG format). Just as a JPG discards redundant pixels while maintaining image quality, this method discards unnecessary global grid data, retaining only the critical local information needed to reconstruct the optimal control signal.

This approach leads to idiosyncratic controllers. Rather than a generic rule, each controller learns a role specific to its location. For instance, a controller at the end of a long feeder line may learn to prioritize voltage support, while one near a substation might focus on bulk power provision. Furthermore, by relying on exogenous variables rather than endogenous ones (like local voltage, which the controller itself influences), the system avoids the feedback instability and "chatter" common in traditional decentralized controls.

Experimental Results

The framework was validated using real-world data from the Pecan Street Dataset (126 residences in Austin, Texas) and tested on realistic models like the 129-Node Feeder and the IEEE 13 Node Test Feeder.

  • Near-Optimal Performance: The decentralized policies achieved results within 0.15% of the centralized OPF solution on average, with a worst-case deviation of only 1.6%.
  • Voltage Regulation: The learned controllers proved highly effective at flattening voltage profiles across the entire feeder, successfully collaborating with legacy hardware like Load Tap Changers (LTCs).
  • Constraint Satisfaction: The models naturally learned to respect hardware capacity limits, with violations occurring in less than 0.00025% of tested cases.
Comparison of network voltage ranges over a 24-hour period. The progression from (a) to (d) illustrates how different control strategies, culminating in the proposed method, successfully regulate and flatten voltage profiles.

Conclusion and Future Directions

This data-driven approach provides a scalable pathway for Distribution System Operators (DSOs) to manage DERs without requiring massive communication infrastructure upgrades. While the current "feedforward" nature of the controllers makes them susceptible to sudden topology changes (like a line break), they offer a robust foundation for modern grid management. Future work aims to explore hybrid models that combine this machine learning approach with minimal feedback loops to handle unexpected grid disturbances and incremental model updates as the grid evolves.

Summary updated

Abstract

The implementation of optimal power flow (OPF) methods to perform voltage and power flow regulation in electric networks is generally believed to require extensive communication. We consider distribution systems with multiple controllable Distributed Energy Resources (DERs) and present a data-driven approach to learn control policies for each DER to reconstruct and mimic the solution to a centralized OPF problem from solely locally available information. Collectively, all local controllers closely match the centralized OPF solution, providing near optimal performance and satisfaction of system constraints. A rate distortion framework enables the analysis of how well the resulting fully decentralized control policies are able to reconstruct the OPF solution. The methodology provides a natural extension to decide what nodes a DER should communicate with to improve the reconstruction of its individual policy. The method is applied on both single- and three-phase test feeder networks using data from real loads and distributed generators, focusing on DERs that do not exhibit inter-temporal dependencies. It provides a framework for Distribution System Operators to efficiently plan and operate the contributions of DERs to achieve Distributed Energy Services in distribution networks.
BibTeX
@article{dobbe2020data,
  author = {Roel Dobbe and Oscar Sondermeijer and David Fridovich-Keil and Daniel Arnold and Duncan S. Callaway and Claire J. Tomlin},
  title = {{Data-Driven Decentralized Optimal Power Flow}},
  journal = {IEEE Trans. Smart Grid},
  year = {2020},
  doi = {10.1109/tsg.2019.2935711},
}
Figure