Online Feedback Optimization (OFO)

OFO treats the physical system as a black box. Instead of relying on an offline system model (like the power flow Jacobian), it applies gradient ascent directly to a cost function measured in real-time. The algorithm perturbs control inputs, observes the resulting change in the measured objective (e.g., reduced losses), and steps toward the optimum.

• Key Distinction: Unlike Model Predictive Control (MPC), no state-space representation is required.
• Benefit: Eliminates errors from model mismatch and parameter drift over time.

The core algorithmic loop estimates the gradient of the objective function with respect to the control inputs using live sensor telemetry. Techniques like extremum seeking or simultaneous perturbation stochastic approximation (SPSA) are employed to isolate the effect of control changes from ambient grid noise.

• Mechanism: Injects a small, continuous probing signal (dither) to measure the local sensitivity.
• Output: A data-driven descent direction that guides capacitor banks and tap changers toward the loss minimum.

OFO operates on a timescale separation principle. It assumes the physical system reaches a quasi-steady state faster than the optimization loop iterates. The algorithm waits for transient dynamics to settle after a control action before taking the next measurement.

• Constraint: The optimization cycle time (seconds to minutes) must be slower than the grid's electrical time constants.
• Stability: This deliberate slowness guarantees that the optimizer does not interact with fast electromechanical oscillations.

While the objective is maximized model-free, hard engineering limits (like ANSI C84.1 voltage bounds) must be strictly respected. OFO enforces these by projecting the gradient step onto a feasible set.

• Method: If a calculated step would violate a voltage limit, the optimizer clips or nullifies that specific control action.
• Safety Layer: A fast overrule logic monitors raw measurements and blocks commands that risk constraint violation, independent of the optimization loop.

Classical optimization uses a model to predict the future state. OFO uses instantaneous feedback as a substitute. By measuring the actual system response to a control change, the algorithm implicitly accounts for unmodeled loads, line impedance errors, and topology changes.

• Adaptation: Automatically tracks the optimal point as solar generation and load mix vary throughout the day.
• Resilience: Remains effective even if the Distribution Management System (DMS) topology model is outdated.

OFO typically acts as a supervisory outer loop, sending updated setpoints to existing local controllers rather than replacing them. It commands the voltage setpoint of a Load Tap Changer (LTC) or the reactive power bias of a capacitor bank controller.

• Architecture: The OFO engine resides in the Distribution Management System (DMS) or an edge gateway.
• Fallback: If communication fails, local controllers revert to their default droop or time-based settings, ensuring no loss of basic voltage regulation.

An advanced control methodology that solves a finite-horizon optimization problem at each time step using a dynamic system model to predict future states. Key distinction from OFO: MPC requires an explicit mathematical model of the plant, while OFO relies purely on real-time measurement feedback.

• Uses a receding horizon to compute optimal control sequences
• Handles multi-variable systems with explicit constraints
• Computationally intensive compared to gradient-based OFO
• Often used as the benchmark against which OFO performance is measured

A linearized mathematical construct, often derived from the power flow Jacobian, that quantifies the incremental change in node voltages resulting from a unit change in reactive power injection or tap position. In OFO, the sensitivity matrix is estimated online from measurement perturbations rather than computed from an offline model.

• Maps control inputs to voltage outputs: ΔV = S · Δu
• Enables gradient computation: ∂Loss/∂u = Sᵀ · ∂Loss/∂V
• OFO estimates S implicitly through iterative perturbation and observation
• Eliminates dependency on accurate network topology and impedance data

A model-free real-time optimization technique that drives a system to the extremum of an unknown objective function by injecting a periodic dither signal and observing the resulting output. OFO is a direct descendant of ESC adapted for discrete-time distribution grid control.

• Uses sinusoidal perturbation to estimate the gradient
• Convergence rate depends on dither frequency and amplitude
• OFO replaces continuous dither with discrete gradient steps
• Originally developed for aerospace and automotive engine optimization

A model-free AI approach where an agent learns an optimal control policy for voltage regulation by interacting with a grid simulation environment. Contrast with OFO: DRL requires extensive offline training on a simulator, while OFO adapts online without pre-training.

• Learns a policy π(s) → a mapping states to actions
• Requires a reward function encoding voltage limits and loss minimization
• OFO is simpler to deploy but lacks DRL's ability to learn long-horizon strategies
• Both methods bypass the need for explicit grid models at runtime

A demand-side management technique that intentionally lowers service voltage to the lower bound of the allowable ANSI C84.1 range to reduce energy consumption. OFO is the ideal real-time control engine for achieving CVR targets without violating voltage constraints.

• Typical voltage reduction: 2-4% below nominal
• Energy savings quantified by the CVR factor (CVRf)
• OFO continuously adjusts reactive power devices to maintain the lower voltage band
• Eliminates the need for pre-computed CVR schedules that cannot adapt to changing load

A deliberate hysteresis zone around a control setpoint within which no corrective action is taken, preventing excessive wear on mechanical equipment. OFO cost functions explicitly penalize unnecessary control actions to balance optimization gains against equipment longevity.

• Typical voltage deadband: ±0.5% to ±1.0% of nominal
• Tap change minimization extends LTC maintenance intervals by 30-50%
• OFO incorporates a regularization term: Loss + λ · (Δu)²
• The penalty weight λ is tuned based on equipment wear cost models