PaperlayerView exact revision on arXiv

Source-linked AI summary

Safety Screening for Voltage Control in Active Distribution Grids via Distributionally Robust Conformal Screening

Sarra Bouchkati, Petros Ellinas, Adriana Geisler, Steffen Kortmann, Johanna Vorwerk, Spyros Chatzivasiliadis, Andreas Ulbig

arXiv:2608.30889v1eess.SYcs.AIcs.LG

TL;DR

New voltage-control policies must be screened before physical deployment despite imperfect simulations and historical data generated under different policies. DR-CSS combines nominal simulation with distributionally robust conformal screening to bound these mismatches, and experiments report target coverage, narrower intervals, and detection of all unsafe gradual-deployment trajectories.

  • Problem

    New policies can induce operating states and interactions that historical data from existing policies and imperfect simulations may not represent, complicating pre-deployment voltage-safety assessment.

  • Method

    DR-CSS uses nominal voltage-trajectory simulations, historical simulation-to-reality error scores, and robust interval enlargement for policy-induced changes in error behavior.

  • Results

    DR-CSS reaches target statistical coverage, reduces average interval width by up to 69%, and flags every unsafe voltage trajectory in the gradual-deployment experiment.

  • Takeaways & Limitations

    Context-aware calibration and gradual policy replacement reduce conservativeness while preserving coverage during staged deployment.

Abstract

from arXiv · show

Deploying a new control policy for voltage control in active distribution grids requires evidence that physical limits will be satisfied before the policy is tested on the physical grid. This assessment is difficult for two reasons. First, simulations cannot capture every disturbance, modeling error, and device interaction present in the real grid. Second, historical measurements reflect operation under existing control policies, whereas a new policy may drive the grid into different operating conditions. To address these challenges, we propose Distributionally Robust Conformal Safety Screening (DR-CSS), a policy-agnostic framework for pre-deployment, scenario-by-scenario screening of a new control policy using historical data and a nominal simulator. For each new scenario, the simulator predicts a future voltage trajectory for the whole grid; DR-CSS then constructs a conformal safety interval around this prediction using historical simulation-to-reality errors. The interval is further enlarged to account for closed-loop changes induced by the deployment of the new policy and its interactions with the remaining controllers. To the best of our knowledge, DR-CSS is the first framework in power systems to combine historical data from an existing control policy with an imperfect simulator for pre-deployment safety screening of a new policy. Experiments on the IEEE 33-bus and IEEE 141-bus systems evaluate the deployment of learning-based voltage control policies and show that DR-CSS identifies all unsafe test scenarios. To reduce unnecessary warnings on safe scenarios, we adapt the safety intervals to different operating conditions and gradually introduce new policies with recalibration after each stage. These extensions increase the informational value of the safety screening and support safer deployment decisions in active distribution grids.

I. INTRODUCTION

Deploying new voltage-control policies requires pre-deployment safety evidence despite imperfect models and historical data collected under existing policies. DR-CSS addresses this off-policy setting with robust, scenario-level conformal screening and practical deployment extensions.

  • Motivation: New policies can alter interactions among inverters, the grid, and remaining controllers, creating a pre-deployment need to assess voltage safety without testing unsafe conditions physically.Historical data may not represent the states, interactions, and control actions induced by an untested policy.
  • Motivation: Imperfect grid models and incomplete observability limit both simulation-based validation and analytical verification of voltage-control policies.Model deviations can prevent analytical guarantees from transferring to physical deployment.
  • DR-CSS: The paper formulates control-policy replacement as an off-policy conformal prediction problem for grid-wide voltage-safety screening.Existing-policy data are used to assess a new policy before deployment.
  • DR-CSS: DR-CSS uses historical simulation-to-reality errors and an imperfect nominal simulator to construct policy-agnostic safety intervals for each new operating scenario.The interval is enlarged to account for bounded changes in prediction-error behavior after policy deployment.
  • Extensions: Context-aware calibration adapts intervals to operating conditions, while gradual policy replacement recalibrates screening after each deployment stage.Staged replacement reduces the change that must be handled when several inverters receive a new policy.

II. SAFETY-SCREENING PROBLEM FORMULATION

The paper models voltage-control deployment as a closed-loop, scenario-level safety-screening problem over future bus-voltage trajectories. It specifies the physical dynamics, local control observations, admissible voltage constraints, and the assumption that the policy is already given.

  • II. SAFETY-SCREENING PROBLEM FORMULATION: The formulation assesses a future trajectory after observing the grid state through time H, with safety evaluated across the look-ahead horizon.The scenario-level event concerns whether all future bus voltages remain admissible.
  • A. Distributed Voltage-Control Setting: The grid contains buses, controllable inverter buses, and load buses; inverter commands are reactive powers, while operating conditions include loads and active-power injections.Bus-voltage magnitudes are constrained between Vmin and Vmax.
  • A. Distributed Voltage-Control Setting: The Volt-Var objective minimizes reactive-power use while requiring every bus voltage to remain within admissible limits and inverter reactive power within capability bounds.The voltage equations use an AC power-flow mapping.
  • A. Distributed Voltage-Control Setting: The study assumes a control policy is already specified and evaluates its safe deployment using historical operating data and a nominal simulator rather than designing the policy itself.The simulator may be an AC power-flow model, reduced-order system, or learned surrogate.
  • A. Distributed Voltage-Control Setting: Each inverter may use local or neighboring measurements, so its policy need not observe the full physical system state.Available measurements can include voltage magnitudes and other inputs such as injections or forecasts.
  • A. Distributed Voltage-Control Setting: The joint policy comprises all local inverter policies, each mapping its observation to a reactive-power setpoint.Policies may be rule-based, optimization-based, or learning-based.
  • A. Distributed Voltage-Control Setting: Closed-loop evolution depends on the physical state, policy actions, grid and device interactions, and exogenous effects such as load changes, PV generation, forecast errors, and disturbances.Future voltages therefore depend jointly on operating conditions, controller actions, and physical response.

B. Safety Screening Goal

The paper formulates pre-deployment safety screening for replacing baseline control policies using historical records, future-condition forecasts, a candidate policy, and a nominal simulator. The goal is to flag scenarios whose candidate-policy voltage behavior is uncertain or unsafe without testing unsafe conditions physically.

  • A deployed candidate policy can alter interactions among inverters, the grid, and remaining controllers, changing closed-loop operating behavior.
  • The tested scenario combines observed grid history, forecasts of future uncontrollable conditions, and the deployed candidate policy.
  • The nominal simulator produces a predicted future voltage trajectory for each scenario, but simulation-to-reality discrepancies arise from model and forecast imperfections.
  • The screening objective is a data-driven rule that assesses candidate-policy scenario safety and flags scenarios that are uncertain or unsafe.

III. DISTRIBUTIONALLY ROBUST CONFORMAL SAFETY SCREENING

DR-CSS compares baseline simulations with realized trajectories, accounts for policy-induced changes, and enlarges the nominal voltage prediction band before making scenario-level decisions. It is an offline screening procedure, not a controller.

  • DR-CSS compares simulated and realized baseline trajectories to quantify simulation-to-reality error, then compares baseline and candidate simulations to capture policy-induced shift.
  • The workflow enlarges the nominal voltage prediction band for both simulation-to-reality error and control-policy replacement effects.
  • DR-CSS applies the same decision rule independently to each scenario and returns warnings for scenarios requiring further testing, policy revision, or exclusion.
  • For each operating scenario, observed history, future uncontrollable-injection forecasts, and the evaluated policy are passed to a nominal simulator that produces one future voltage trajectory.
  • Using the simulator trajectory directly would treat the simulator as exact, so DR-CSS instead constructs a data-calibrated uncertainty band around it.

B. Voltage Error Scores and Conformal Calibration

DR-CSS calibrates scenario-level voltage uncertainty from historical nominal simulations and realized voltage profiles. Split conformal prediction converts these errors into a finite-sample threshold under an exchangeability condition.

  • Historical nominal simulations are paired with realized physical-grid voltage profiles, and each discrepancy is summarized by a scalar nonconformity score.
  • The score captures the largest standardized voltage-prediction error across all buses and future time steps for each calibration scenario.
  • The calibration score is dimensionless and uses fixed bus-specific residual scales and future-time weights; experiments use uniform time weights, γt = 1.
  • The sorted calibration scores provide a conformal threshold at target miscoverage α, with α = 0.05 corresponding to a nominal 95% trajectory-level coverage target.
  • The resulting voltage interval is reliable when calibration and future errors are exchangeable, but policy replacement can change the closed-loop conditions underlying those errors.

C. Robust Enlargement for Policy-Induced Error Changes

DR-CSS addresses policy-induced changes by inflating the conformal score quantile, selecting the robustness budget on candidate-policy tuning data, and applying the resulting band to complete voltage trajectories. Larger budgets improve protection but widen intervals and can increase conservative warnings.

  • Policy replacement can shift or thicken the scalar simulation-error distribution, making baseline calibration scores less representative of candidate-policy operation.
  • The distributionally robust step uses a higher conformal score quantile while leaving the nominal simulated trajectory unchanged.
  • Larger robustness budgets produce wider, more conservative voltage intervals, whereas small budgets may insufficiently protect against policy-induced error changes.
  • The robustness budget is selected on an independent tuning set generated under the candidate deployed policy, using a Wilson lower confidence bound to avoid optimistic empirical coverage.
  • Monotonicity enables bisection to find the smallest feasible robustness budget, limiting unnecessary interval enlargement and overly conservative warnings.
  • DR-CSS accepts a candidate deployment only when the complete robust voltage band stays within admissible limits across all buses and future time steps.

D. Scenario-Level Accept/Warning Decision

DR-CSS makes a scenario-level deployment decision by checking whether the full robust voltage interval stays within admissible limits across all buses and future times.

  • The final interval is formed after selecting the robustness budget and defining the corresponding robust conformal threshold.
  • DR-CSS accepts a candidate only when the final robust interval lies within voltage limits for every bus and forecast time.Otherwise, it issues a warning for the scenario.
  • Under the stated assumptions, the robust interval provides prescribed joint statistical coverage for the post-deployment grid-wide voltage trajectory.A warning indicates insufficient evidence from the simulator, calibration data, and robustness budget, not that a violation will occur.

E. Context-Aware Calibration

Context-aware calibration addresses operating-condition-dependent policy shifts by assigning scenarios to net-active-power groups and calibrating a separate robust threshold for each group.

  • Policy-induced distribution shift varies with operating conditions because load demand and PV generation affect inverter actions and voltage responses.A single global threshold can therefore be overly conservative in some conditions and insufficient in others.
  • A separate robust threshold is calibrated within each operating group and selected at test time using the observed condition.The accept/warning rule remains unchanged.
  • Scenarios are grouped by empirical quantiles of terminal-history net active power, which measures the balance between load demand and PV generation.Negative values indicate PV-dominated operation, while larger values indicate higher net demand.

IV. EXPERIMENTAL RESULTS

The experiments evaluate DR-CSS on IEEE 33-bus and IEEE 141-bus systems using droop control as the baseline and IDDPG-based learning control as the candidate policy.

  • DR-CSS is evaluated on IEEE 33-bus and IEEE 141-bus systems in single-control, context-aware, and gradual replacement experiments.
  • The experiments use MAPDN with publicly sourced PV and residential-load profiles resampled at three-minute resolution.Each daily profile defines an operating scenario spanning varied generation and demand conditions.
  • The admissible voltage range is [0.95, 1.05] p.u., with reactive-power capability constrained by inverter apparent-power ratings.
  • Trajectory windows contain 21 history steps and a 10-step prediction horizon, corresponding to 63 minutes of history and a 30-minute look-ahead.
  • The baseline uses piecewise-linear local droop control, while the candidate uses individual IDDPG actors assigned to inverter observation zones; remaining inverters retain droop control.

2) Benchmark Methods:

The benchmark comparison tests standard conformal prediction and Weighted MA-COPP against DR-CSS for off-policy voltage-trajectory screening, including spatial and load-condition interval behavior.

  • Benchmark Methods: The comparison uses the same simulator and a conformal miscoverage level of αCP = 0.05, corresponding to 95% nominal coverage.
  • Benchmark Methods: Standard CP calibrates split-conformal intervals on all-droop data, whereas Weighted MA-COPP applies importance-weighted conformal prediction in the off-policy setting.
  • Overall Performance: 100% of Weighted MA-COPP critical values are infinite, producing unbounded intervals; capping them removes the original coverage guarantee.
  • Overall Performance: DR-CSS restores nominal coverage without unbounded intervals, with average widths only 3.0% and 5.0% above Standard CP on IEEE 33-bus and IEEE 141-bus systems.Compared with capped Weighted MA-COPP, DR-CSS reduces average width by approximately 69% and 63%, respectively.
  • Overall Performance: Prediction intervals widen toward feeder ends, where control actions have greater influence on voltage and policy-induced distribution shifts are larger.

2) Context-Aware Calibration:

Context-aware and gradual calibration address operating-regime dependence and the larger distribution shifts caused by replacing multiple control policies. Gradual deployment preserves coverage while reducing conservativeness and detects every unsafe trajectory in the reported experiment.

  • Context-Aware Calibration: Loadwise DR-CSS attains the 1 −α = 0.95 coverage target in every operating-point bin, unlike standard CP and uniform DR-CSS.Standard CP and uniform DR-CSS fall below target outside the two most PV-dominated bins, with the largest undercoverage in bin 4.
  • Gradual Deployment: Replacing several control policies simultaneously requires a stronger robustness correction and can increase warning decisions.The gradual strategy instead introduces policies sequentially, collecting new data and recalibrating after each replacement.
  • Gradual Deployment: Approximately 59% lower robustness correction and 36% lower conservativeness are achieved by gradual replacement than by abrupt replacement under the same final 3RL deployment.Both strategies exceed the nominal 95% coverage level.
  • Gradual Deployment: Gradual replacement yields a 5.62% target miss rate versus 7.36% for abrupt transition, closer to the nominal 5% miscoverage level.Sequential data collection and recalibration reduce the effective distribution shift and tighten prediction intervals while maintaining coverage above target.
  • Warning Performance: Under gradual policy shift, trajectory-level recall is 1.0000, detecting every rollout with at least one voltage violation.The method flags 77.10% of rollouts, with trajectory-level precision of 0.5447.
  • Warning Performance: The screening prioritizes avoiding missed voltage violations at the cost of conservative alarms.At point level, recall is 0.9965 while 9.65% of checks are flagged and precision is 0.1610.

V. CONCLUSION

The paper proposes DR-CSS to screen new voltage-control policies using historical data and imperfect simulations while accounting for simulation and policy-induced distribution shifts. Its evaluations support context-aware and gradual deployment strategies, and the paper identifies continuous conformal risk control and broader grid applications as future directions.

  • DR-CSS combines nominal simulations with distributionally robust conformal screening to bound simulation-to-reality and policy-induced distribution shifts.
  • The framework is presented as an efficient safety screening approach for new voltage-control policies under imperfect simulations and historical data collected under another policy.
  • Future refinement may use continuous conformal risk control to optimize expected physical severity or duration of minor overvoltages.
Loading 2608.30889v1…