Model Predictive Control (MPC)

MPC solves an open-loop optimal control problem over a finite prediction horizon at each sampling instant, but only the first control action is implemented. The horizon then shifts forward one step and the optimization repeats. This receding horizon strategy provides inherent feedback to counteract model inaccuracies and unmeasured disturbances.

• Prediction Horizon: Typically 15-60 minutes for distribution VVO, capturing load and solar variability
• Control Horizon: Shorter than prediction horizon; only the first step is executed
• Computational Cycle: Must complete within the sampling interval (e.g., 30 seconds to 5 minutes)
• Contrasts with open-loop day-ahead scheduling which cannot adapt to real-time deviations

MPC explicitly incorporates a state-space model of the distribution grid, capturing the dynamic coupling between control inputs (tap positions, capacitor states, inverter reactive power) and outputs (node voltages, line flows). This model predicts the trajectory of system states over the horizon.

• Linear Time-Varying (LTV) Models: Derived from the power flow Jacobian at the current operating point for computational tractability
• Nonlinear Models: Full AC power flow equations used when accuracy demands it, solved via sequential quadratic programming
• Disturbance Modeling: Load forecasts and PV generation predictions are treated as known exogenous inputs
• State estimation provides the initial condition for each MPC iteration

A defining advantage of MPC is its systematic ability to enforce hard constraints on both control inputs and state variables. For Volt-VAR control, this means voltage limits (ANSI C84.1 Range A) and equipment operational limits are explicitly respected in the optimization.

• State Constraints: Maintain all bus voltages within [0.95, 1.05] per unit
• Input Constraints: Tap changer range limits (e.g., ±10% in 16 or 32 steps), capacitor bank discrete switching states
• Rate Constraints: Limit tap changes per hour to preserve Load Tap Changer (LTC) lifespan
• Security Constraints: Thermal limits on lines and transformers prevent overloads
• Constraint violations are anticipated and preempted, not just corrected after detection

The MPC controller minimizes a weighted sum cost function that balances competing operational objectives. This scalarization allows the utility to tune the trade-off between conservation voltage reduction, loss minimization, and equipment preservation.

• CVR Objective: Minimize total active power consumption by driving voltages to the lower ANSI band
• Loss Minimization: Reduce I²R losses by optimizing reactive power flows and voltage profiles
• Control Effort Penalty: Weight on tap changes and capacitor switching to minimize mechanical wear
• Voltage Deviation Penalty: Soft constraint formulation using quadratic penalty terms for voltage limit violations
• Demand Response Coordination: Integrate cost signals from wholesale markets or peak demand charges

MPC combines feedforward action based on predicted disturbances (cloud cover, load ramps) with feedback correction from real-time measurements. This dual mechanism enables proactive voltage regulation before violations occur while correcting for model mismatch.

• Feedforward Path: Anticipates voltage rise from impending PV ramp events and pre-positions reactive power resources
• Feedback Path: State estimator updates correct the internal model at each time step, closing the loop
• Disturbance Rejection: Faster response to voltage sags from motor starts or feeder faults than purely reactive control
• Contrasts with rule-based VVO which reacts only after a threshold crossing is detected

Volt-VAR MPC inherently involves discrete decision variables (capacitor bank on/off, LTC tap positions) alongside continuous variables (inverter reactive power setpoints). This leads to a Mixed-Integer Nonlinear Programming (MINLP) or Mixed-Integer Quadratic Programming (MIQP) formulation.

• Branch-and-Bound: Systematic tree search to find globally optimal discrete decisions
• Heuristic Decomposition: Separate discrete and continuous subproblems to achieve real-time performance
• Convex Relaxation: Relax integer constraints to solve a convex problem, then round to nearest feasible integer
• Embedded Solver: Lightweight QP solvers deployed on edge computing platforms within substation gateways
• Computational tractability is the primary engineering challenge for field deployment

The mathematical backbone of MPC for Volt-VAR control. MINLP formulates the optimization problem by handling discrete control variables—such as capacitor bank steps (on/off) and load tap changer positions (integer ratios)—alongside continuous variables like voltage magnitudes and reactive power injections. The MPC solver must navigate this non-convex landscape at each time step, balancing global optimality against the computational speed required for real-time grid operations.

A linearized construct derived from the power flow Jacobian that quantifies how node voltages react to control changes. In MPC, the sensitivity matrix serves as the internal prediction model, mapping a unit change in reactive power injection or tap position to its incremental voltage impact across the feeder. This allows the optimizer to rapidly evaluate thousands of candidate control trajectories without executing a full nonlinear power flow for each prediction step, dramatically accelerating the solution time.

The algorithmic engine that provides MPC with its initial system state. The DSE processes redundant, noisy, and asynchronous SCADA and AMI measurements to compute the most probable voltage and current phasors for every node. Without an accurate state estimate, the MPC's internal model diverges from physical reality, leading to suboptimal or destabilizing control actions. The DSE acts as the bridge between raw telemetry and the optimizer's digital twin.

A primary actuator controlled by MPC for voltage regulation. The LTC adjusts a transformer's turns ratio under load via mechanical or solid-state switching. MPC must explicitly model tap change minimization as a penalty in its cost function to prevent excessive hunting and mechanical wear. The optimizer predicts the long-term voltage trajectory and dispatches tap commands only when the benefit outweighs the accumulated maintenance cost over the prediction horizon.

A competing real-time control paradigm that bypasses the need for a precise offline model. Unlike MPC, which relies on an internal system model for prediction, OFO applies gradient steps computed from live measurements to iteratively drive the grid toward optimality. It is often compared to MPC in scenarios where maintaining an accurate digital twin is cost-prohibitive, though MPC generally offers superior constraint handling and anticipatory action.

A primary objective often embedded within the MPC cost function. CVR intentionally lowers service voltage to the lower bound of ANSI C84.1 limits to reduce energy consumption. The MPC controller predicts the CVR factor (CVRf)—the ratio of demand reduction to voltage reduction—across the horizon and optimizes the voltage profile to maximize energy savings while ensuring the last customer on the feeder remains above the critical undervoltage threshold.