Grid-Level Optimization in Electric Grids

• Grid-level optimization is the systematic control of grid components to optimize reactive power flows under operational and economic constraints.
• It employs online feedback optimization to coordinate inverter settings, addressing limitations of local droop control for enhanced voltage regulation.
• Experimental results indicate that OFO increases active power export by ~9-10% over droop methods, deferring costly grid reinforcements.

Grid-level optimization concerns the systematic determination of optimal control actions, setpoints, or device parameters across the entire electric grid—or localized subgrids—to satisfy system objectives under operational, technical, and economic constraints. In power systems, this often entails maximizing renewable (e.g., PV) hosting capacity or active power transfers subject to voltage, thermal, and device limitations. In low-voltage (LV) distribution networks, a critical challenge is enhancing grid capacity through advanced reactive-power (Volt/VAr) control to delay or avoid physical reinforcement. Recent research highlights the limitations of local control laws and demonstrates the potential of coordinated, model-lite feedback optimization to unlock the full capability of distributed inverters and DERs, allowing substantially higher active power throughput at minimal cost (Matt et al., 2023).

1. AC Power-Flow Model and Grid Constraints

The underlying physical and operational constraints of grid-level optimization are governed by the steady-state AC power flow equations. For a network with buses indexed by $i=0$ (slack) to $N$, the complex bus voltages $V$ and admittance matrix $Y$ define the injections: $$S_i = P_i + jQ_i = V_i \sum_{j=0}^N Y_{ij}^* V_j^*,\quad i=0,\dots,N,$$ with

$$P_i = \text{active power} = P_i^{\mathrm{gen}} - P_i^{\mathrm{load}},$$

$$Q_i = \text{reactive power} = q_i + Q_i^{\mathrm{gen}} - Q_i^{\mathrm{load}},$$

where $q_i$ is the controllable inverter reactive power. Operational constraints enforce

$$v_{\min} \leq |V_i| \leq v_{\max},\qquad q_{i,\min} \leq q_i \leq q_{i,\max}$$

for all $i=1,\dots,N$.

The optimization goal is often to maximize aggregate active injection subject to these constraints, which physically translates into supporting higher renewable export without voltage or hardware violations.

2. Formulation of the Reactive Power Coordination Problem

A canonical formalism for grid-level reactive power optimization is the Optimal Reactive Power Flow (ORPF): $N$0 where $N$1 is the voltage magnitude at bus $N$2 as a function of $N$3 (controls) and $N$4 (disturbance: net active and reactive load/gen). $N$5 is the box-constrained inverter capability region. The objective can be adapted to minimize other grid-level costs (e.g., losses, curtailment). At each time, this nonlinear, constrained problem yields the globally optimal allocation of reactive resources to regulate voltages across the network.

In practical deployment, the inverters typically operate with decentralized “droop” controllers: $N$6 with a deadband $N$7 and deadzone linear gains. This achieves only local voltage regulation, neglecting global interactions and reactive sharing effects.

3. Online Feedback Optimization: Algorithmic Principles

Online Feedback Optimization (OFO) is a closed-loop control framework that iteratively drives the inverter setpoints $N$8 towards the solution of (ORPF) using only real-time voltage measurements, basic model information (a voltage sensitivity matrix $N$9), and communication with reactive sources. The controller maintains dual variables $V$0, $V$1 for the lower and upper voltage constraints. The update laws are: $V$2

$V$3

$V$4

$V$5

where $V$6 projects onto the inverter capability set.

At each cycle, inverters measure $V$7, the controller updates duals and new setpoints, and dispatches $V$8 to each inverter. The method tolerates significant model inaccuracies—the only requirement is a rough sensitivity matrix $V$9—and does not require knowledge of non-controllable loads or distributed PV generation. Under reasonable $Y$0, convergence to the ORPF optimizer is guaranteed, and voltage constraints are enforced in steady state.

4. Performance Comparison: Local Droop vs OFO

Quantitative analysis demonstrates the substantial limitations of local droop control:

• Inverters electrically close to the substation contribute little—reactive capacity is not deployed where needed.
• Others deep in the feeder saturate early, failing to compensate downstream voltages.
• Overall, droop delivers sub-optimal voltage regulation and excessive reactive energy usage.
• Machine-learning-tuned droops yield only marginal improvement, but inherit fundamental structural limitations.

By contrast, OFO achieves nearly the full globally optimal improvement provided by ORPF, consistently outperforming all local surrogates. In year-long simulations (European CIGRÉ LV feeder, household models from real data), for high PV-penetration scenarios, the OFO scheme increases steady-state maximum active power export by 9% over droop, and eliminates unnecessary reactive consumption. In a hardware-in-the-loop field experiment, OFO extended the PV hosting limit by 10.5% compared with droop, confirming simulation results (Matt et al., 2023).

5. Experimental Implementation and Scalability

OFO is designed for practical deployment:

• Computation per step is $Y$1; linear updates and box projections are lightweight.
• Scalable: can run centrally or federated (inverter-level or aggregator-level).
• Communication is confined to setpoint exchange with reactive sources—no wide-area measurement is required.
• The only model information is an approximate voltage–reactive sensitivity matrix $Y$2, which can be derived from a nominal network model or measured through system identification.
• Tuning reduces to selection of the (global) dual update gain $Y$3; stability is robust for all $Y$4. No explicit model-based retuning is needed for changes in load, PV, or topology.

Deployment requires modern inverters with voltage sensing/telemetry and a channel (wired/WiFi/PLC) for setpoint dispatch.

6. Implications, Limitations, and Practical Recommendations

The core finding is that reactive power coordination via online, measurement-based optimization can unlock virtually all the “virtual” grid reinforcement latent in a network’s inverter fleet, deferring expensive line/transformer upgrades. OFO achieves this with minimal infrastructure beyond what exists for remote control or advanced metering, with convergence guarantees and resilience to moderate model error.

In comparison to local droop and ML-tuned droop:

• OFO adapts online to time-varying grid conditions and DER deployments, without the need for repeated offline retraining.
• Reactive energy is automatically minimized, and hosting capacity is always maximized.
• The virtual reinforcement effect—an additional ∼10% in hosting capacity—is nearly invariant across feeder types (simulated and real), making OFO an attractive default for utilities upgrading Volt/VAr capability.

Limitations include the need for modest communication and setpoint dispatch infrastructure, and the assumption that voltage magnitude measurements are available at each controllable node.

In summary, grid-level optimization for Volt/VAr enhancement in LV networks requires moving beyond uncoordinated, static control laws. Online Feedback Optimization, with model-agnostic, measurement-driven coordination, represents a scalable, cost-effective, and operationally validated solution to maximize grid utilization and defer costly reinforcements (Matt et al., 2023).