Learning and validating local smart-inverter controls

This tutorial gives a reproducible workflow for learning a local inverter control law from data, then evaluating it with PowerOptLab. The target laws are constant power factor, Volt-VAr, and Volt-Watt.

The central rule is simple: a deployed controller is a local policy, not a free OPF decision. It observes signals at its point of connection (POC), applies a prescribed response, and thereby changes the voltages it will observe next. An OPF can be a useful benchmark or teacher, but its setpoints often depend on non-local information unavailable to a physical inverter.

What PowerOptLab evaluates

PowerOptLab does not yet train a learning model. It provides the physically consistent evaluation stage: an IBR can reference a BMOPFTools control_profile, and that prescribed control remains active in power flow, OPF, and dynamic operating-envelope (DOE) calculations. Fit policy parameters externally, then attach the fitted profile to the IBR.

Keep these three tasks separate:

Choose a deployable policy class first

A constant-power-factor law imposes, per phase,

\[\operatorname{sign}(pf) Q + \tan(\arccos |pf|) P = 0.\]

Positive pf is lagging (absorbing VAr) and negative pf is leading (injecting VAr). It is appropriate for a deliberately fixed operating mode, but it does not provide voltage feedback.

Volt-VAr uses four voltage breakpoints and two reactive limits: it injects VAr at low voltage, has a zero-VAr deadband, and absorbs VAr at high voltage. Volt-Watt uses two high-voltage breakpoints to cap active export. Together they make a closed loop:

local voltage → Volt-VAr / Volt-Watt curve → P/Q injection → feeder voltage → local voltage

A profile normally contains one law. If power_factor and a droop are both present, fixed power factor takes precedence with a warning. Do not fit both and mistake the result for a hybrid controller.

Construct an identification data set

For each inverter and time index, retain local voltage, measured P and Q, available active power, controller mode, rating, terminal/phase, timestamps, curtailment commands, and quality flags. The voltage must be the inverter's POC measurement—not a substation or another bus. Without available PV power, Volt-Watt curtailment cannot be distinguished from low irradiance.

Split training and test data by day, weather event, feeder state, or customer cluster. Randomly splitting adjacent timestamps leaks the same voltage trajectory into both sets and greatly overstates generalisation.

Fit constrained, interpretable curves

For Volt-VAr, fit ordered breakpoints and VAr limits subject to apparent-power capability and grid-code constraints. For Volt-Watt, fit ordered breakpoints and a non-increasing cap in the curtailment region. These constraints are essential: a flexible regressor can fit noise with a non-monotone response that creates positive feedback once it is placed in the network loop.

Use only features that the installed device can measure. If the learner consumes remote voltage, peer injections, or a state estimate, it is a centralized policy and must be evaluated with its communication and delay assumptions.

Replay the learned profile

Translate fitted parameters into a control profile, then bind it to the IBR.

net["control_profile"] = Dict(
    "learned_pv" => Dict(
        "volt_var" => Dict(
            "voltage_reference" => "PN_PER_PHASE",
            "breakpoints" => [207.0, 220.0, 240.0, 258.0],
            "q_limits" => [-0.60, 0.44],
            "q_unit" => "VA_FRACTION",
            "q_ref" => "VAR_MAX",
        ),
        "volt_watt" => Dict(
            "voltage_reference" => "PN_PER_PHASE",
            "breakpoints" => [253.0, 260.0],
            "p_limits" => [0.20, 1.00],
            "p_unit" => "VA_FRACTION",
            "p_ref" => "P_AVAILABLE",
        ),
    ),
)
net["ibr"]["pv17"]["control_profile"] = "learned_pv"

The policy is an equality/control constraint, not a suggestion that the OPF may override. To test it in an active-power DOE, bind the point to its IBR:

cp = ConnectionPoint(id="customer_17", bus="lv17", ibr_id="pv17",
                     export_max=10e3)
r = solve_operating_envelope(net, [cp]; direction=:export)

The DOE changes active capacity only. It retains the IBR topology, apparent/current limits, and Volt-VAr/Volt-Watt equality; it does not invent a separately dispatchable Q source.

Pitfall: units and signs

VA_FRACTION scales a limit by rating; VAR and W are absolute. Volt-VAr limits are ordered [q_absorb ≤ 0, q_inject ≥ 0]. Check the fitted curve at a low-voltage point, a deadband point, and a high-voltage point before feeder runs.

Validate the closed loop

One-step P/Q prediction error is not sufficient. Compare the fitted policy with a mandated/default curve and fixed-PF baseline on held-out feeder scenarios:

1. high-PV/low-demand and low-PV/high-demand cases;
2. source-voltage, measurement, and parameter uncertainty;
3. simultaneous nearby-inverter response, not one-at-a-time replay; and
4. voltage, thermal, neutral, apparent-power, curtailment, and VAr-throughput metrics.

Use voltage-violation frequency and magnitude, total curtailed energy, and constraint margins as primary metrics. A small error near a steep knee can move the closed-loop equilibrium substantially.

Multi-phase pitfalls

The voltage reference is physical. PN_PER_PHASE is phase-to-neutral, PG_PER_PHASE phase-to-ground, and PP_PER_PHASE phase-to-phase; each can be averaged across phases. In a four-wire feeder with neutral displacement, phase-to-ground and phase-to-neutral are not interchangeable, while averaging can hide the phase that needs support.

Volt-VAr/Volt-Watt droop is supported for SINGLE_PHASE and FOUR_LEG IBRs. A THREE_LEG IBR lacks the degrees of freedom for the per-phase droop model and falls back to box bounds with a warning. Do not validate a per-phase learned controller with that topology.

A defensible study protocol

1. State whether the task is identification, local-policy design, or OPF imitation.
2. Declare deployment-time measurements and actuation signals.
3. Fit a constrained, grid-code-compliant policy with event-wise data splits.
4. Replay the exact profile in the nonlinear feeder model.
5. Validate simultaneous closed-loop behaviour on held-out, stressed, and unbalanced cases where relevant.
6. Report the centralized OPF only as a benchmark unless its information and communications are actually deployable.

This separation between learned data behaviour, network feasibility, and local device information is what turns an attractive fitted curve into a credible smart-inverter control study.