Additive Volt/Var Policies That Operators Can Read and Edit

Rooftop solar and electric-vehicle charging now make voltages on local distribution feeders swing within seconds. The inverters attached to solar panels can push or absorb reactive power to steer those voltages back, and reinforcement learning (a controller that learns by trial and error in simulation) is very good at deciding how much. Utilities still hesitate to deploy these controllers. The learned policies are opaque, and the most reliable results in the field come from forcing the policy into a constrained shape that guarantees stability. We propose to give every inverter a controller built the way interpretable clinical risk models are built. Its action is a sum of one-variable curves: one curve answers "how hard do I react to my own voltage error," and the others answer "how much do I pre-adjust given my solar output, local load, and time of day." Each curve can be plotted, checked against grid codes, and edited by an engineer. The voltage curve is monotone by construction, which keeps the known stability guarantee in place. The other curves are reluctant: they stay at zero unless they measurably improve control.

The idea comes from a simple observation. The existing stability-guaranteed controllers are already interpretable, but only because they react to local voltage alone. The real question is whether richer inputs help, and whether those inputs can be added without losing either the stability proof or the ability to audit the controller. An additive structure is what allows both. The derivative of the action with respect to local voltage then depends only on the voltage curve, so monotonicity in voltage holds no matter what the other inputs are doing. We test whether this buys faster recovery, lower losses, and edits that survive a change of season, relative to strong black-box and voltage-only controllers.

Core Claim

Each inverter $i$ uses the policy

$$q_i(t) = -\phi_i\big(v_i(t)-1\big) + \sum_{k} g_{ik}\big(z_{ik}(t)\big),$$

where $q_i$ is reactive power injection and $v_i$ is local voltage in per-unit. The term $\phi_i$ is a non-decreasing curve through the origin. The $z_{ik}$ are inputs the controller does not itself drive, such as local PV active output, local load, and time of day, and each $g_{ik}$ is a one-dimensional learned curve. Because $\partial q_i/\partial v_i = -\phi_i'$ regardless of $z$, the strict-monotonicity condition used by Stability Constrained Reinforcement Learning for Real-Time Voltage Control still holds for the voltage-feedback part. The exogenous terms act as a feedforward shift of the operating point. For fixed exogenous inputs this is equivalent to a change in net injection, so the equilibrium moves but its stability is preserved. Under time-varying inputs we expect an input-to-state style bound, and we will check it empirically rather than assume it.

We make three claims. First, the feedforward curves reduce transient recovery time and reactive losses relative to voltage-only monotone policies. Second, the policy remains exactly readable as a handful of plots per inverter. Third, when performance degrades across seasons, the additive structure localizes which curve went wrong, so that one curve can be refit or clipped by an operator without retraining and without touching the stability guarantee.

How the Pieces Are Borrowed

The clinical-interpretability literature supplies the model form. Neural Additive Models fit one small network per input and sum their outputs. They show on MIMIC-II mortality that this gives up little accuracy (AUC 0.830 against 0.844 for XGBoost), and that individual curves can be zeroed to remove an unwanted effect without retraining. We reuse their exp-centered units, $h(x)=f(e^{w}(x-b))$. The $e^{w}$ scaling is already positive, so pairing it with a monotone activation and non-negative output weights makes $\phi_i$ monotone at almost no cost. The original NAM work has no monotonicity constraints and makes no claim about stability under closed-loop control; we add both.

A reluctant additive model framework for interpretable nonlinear individualized treatment rules supplies the complexity discipline. Its rule starts from a sparse linear fit and admits nonlinear spline terms only if they explain residual signal beyond a heavier penalty. In our setting the linear starting point is a classical droop curve, the familiar IEEE 1547 style Volt/Var slope. The nonlinear part of $\phi_i$ and every $g_{ik}$ must earn its place under a group penalty, with term selection by held-out control cost. RAITR selects terms for treatment rules from static data. We select them inside an actor-critic training loop, and that loop is a real source of risk, discussed below.

The power-systems side supplies the guarantee and the safety layer. The monotone policies of Stable-DDPG and TASRL motivate the constraint. TASRL's control-barrier-function quadratic program, a small optimization that minimally adjusts actions to keep voltages inside limits, will wrap our policy unchanged, so steady-state optimality and hard limits come from that layer. Our contribution is the transient policy inside it.

The closest cross-domain work occupies different ground. Explaining reinforcement learning policies for power grid operations gives post-hoc counterfactual explanations for transmission topology switching and reports that they are noise-sensitive. Our explanations are the policy itself, applied to distribution Volt/Var. Large Language Model as An Operator offers readable reasoning but works at 15-minute steps on taps and capacitors, without physical guarantees. A Neuro-Symbolic Approach to Multi-Agent RL produces logic rules for chip power budgets and has no AC power flow. The opacity complaints in Multi-Agent Reinforcement Learning for Active Voltage Control on Power Distribution Networks motivate the work. Its finding that high-capacity multi-agent controllers are needed at scale is the tension this proposal tests.

Research Program

We will build the additive monotone policy, train it with DDPG inside the TASRL safety layer on pandapower feeders, and analyze it in three ways. The first is a capacity analysis: how much of the gap between voltage-only monotone control and an unconstrained multi-agent controller the feedforward curves recover. The second is an editing analysis: operators' typical interventions, such as capping reactive output at low solar or flattening a time-of-day term, are applied directly to curves and evaluated for safety and cost. The third is a shift analysis: training on summer profiles, measuring which curves change most when refit on winter data, and testing whether refitting only those curves recovers performance.

One technical risk is that the reluctance penalty interacts badly with critic noise. Nonlinear terms might be admitted or pruned because of bootstrapping error rather than genuine benefit. We will report term selection stability across seeds as a primary diagnostic.

Baselines

The naive reference points are no control and a tuned piecewise-linear droop. The tuned droop uses a per-inverter slope and deadband, fit by grid search under the same simulator budget as the learned methods. If the reluctance penalty ends up admitting no curves, our method should collapse to exactly this, which makes it the sanity floor.

The strongest stability-guaranteed competitor is Stable-DDPG, a voltage-only monotone neural policy. On the SCE 56-bus circuit it reports 100% stability across 500 scenarios, against 30–40% failures for unconstrained DDPG, plus 33% faster recovery than linear policies. Its transient-plus-steady-state successor TASRL reports 12.08 seconds voltage recovery on IEEE 123-bus, 30% faster than stable RL and 77% faster than pure safe gradient flow. We take TASRL as the main transient baseline. Our policy replaces its transient network inside the identical safety layer, with matched environment steps, seeds, and hyperparameter search budget. Beating it means faster recovery or lower transient cost at the same 100% stability over the same 500 scenarios.

For the quasi-static multi-agent setting, the strongest competitor is MA-DELC from Safety Constrained Multi-Agent Reinforcement Learning for Active Voltage Control. It reports a control ratio near 1.0 on 33- and 141-bus and about 0.90 on 322-bus. MADDPG and MATD3 from the MAPDN benchmark reach near-100% controllable rates on smaller feeders. Here we would count matching MA-DELC's control ratio within noise, with lower reactive losses, as success. We do not expect to beat it. The point is that a readable controller matches it.

Two ablations isolate the key ingredient. One drops all $g_{ik}$, giving a voltage-only reluctant monotone policy. The other replaces the additive form with a partially monotone MLP that is monotone in $v_i$ but has unconstrained interactions with $z$. The second ablation separates the effect of additivity from the effect of extra inputs.

Why not just…?

Why not just plot Stable-DDPG's policy? Because it is a function of local voltage only, it is already a readable curve. If the feedforward terms add nothing, the proposal reduces to this. The voltage-only ablation therefore decides the question directly. If it ties the full model on recovery time and losses at 56- and 123-bus, we will report that interpretability was never the bottleneck for voltage-only control, and that richer inputs are not worth adding.

Why not just tune droop curves well? Droop is fully transparent. However, the MAPDN study and Optimal Power Flow Pursuit via Feedback-based Safe Gradient Flow both report that static droop violates the 1.05 p.u. limit under fast PV transients and incurs higher losses. Our tuned-droop baseline gives droop a fair search budget. Because reluctance starts from droop, any gain we report is gain that droop demonstrably could not achieve.

Why not train MADDPG and distill it into a GAM, or explain it with SHAP? Post-hoc attributions approximate the policy rather than being it, as the NAM authors argue. The grid counterfactual study found its explanations unstable under small state noise. A distilled GAM can be compared fairly, though, and it is a real threat. We will fit an explainable boosting model to MADDPG actions and deploy it in closed loop. A distilled model carries no monotonicity guarantee, so we expect stability failures resembling those of unconstrained DDPG. If it is stable and matches our cost, our training-time construction offers little beyond convenience, and we will say so.

Why not a partially monotone MLP with all inputs? It keeps the guarantee and may control better. This is the second ablation above. If it wins by a wide margin, additivity costs too much, and the honest conclusion is that operators must accept interaction terms. We would then test whether adding a few bivariate curves closes the gap.

Proposed Experiments

Transient experiments use pandapower on the SCE 56-bus and IEEE 123-bus feeders, with the 500 random violation scenarios used by Stable-DDPG and TASRL. We report stability rate, recovery time, and transient reactive cost.

Quasi-static and scaling experiments use the MAPDN benchmark on 33-, 141-, and 322-bus networks, with its Portuguese load and Elia solar data. We report control ratio and reactive losses, training on summer and testing on winter. The 322-bus feeder tests the scaling wall: if decentralized additive policies fall well below MA-DELC's roughly 0.90 there, the tension in the literature resolves against readable controllers at large scale.

The editing experiment applies a fixed set of operator-style curve edits on the 141-bus feeder and measures violation rate and cost change with no retraining. Edits to $g$ terms should never break stability, by construction. Edits to $\phi$ that preserve monotonicity should not break it either. The shift experiment refits only the top-$k$ most-shifted curves on winter data and compares the result against full retraining.

The decisive checks are three. The full model must beat the voltage-only ablation on transient cost at 100% stability. It must stay within noise of MA-DELC on 141-bus. And partial refitting must recover most of the winter performance loss. Failing the first check ends the proposal's main claim. Failing the second limits it to small feeders.