Source-linked AI summary
Stochastic Reactive Power Management in Microgrids with Renewables
Vassilis Kekatos, Gang Wang, Antonio J. Conejo, Georgios B. Giannakis
TL;DR
Renewable-rich distribution microgrids face uncertain injections, reverse power flows, and voltage fluctuations that challenge conventional reactive-power management. The paper develops a distribution-free online stochastic controller using power-injection data, dual SOCP information, and reactive-power compensation incentives. Numerical tests show lower costs than myopic deterministic control, successful tracking of system variations, and convergence toward ideal control.
Problem
Distribution microgrids experience uncertain renewable generation and demand, reverse power flows, and voltage fluctuations, while existing reactive-control approaches assume precisely known, unchanged active-power injections.
Method
The paper models active injections as stochastic processes and uses a provably convergent stochastic-approximation controller whose reactive updates derive from the dual SOCP and power-injection data.
Results
12.7$/h benefit over the deterministic scheme is reported in one experiment, while savings reach 45.6 $/h for η = 4 in another comparison.
Takeaways & Limitations
The stochastic framework adapts dynamically to microgrid operation data without distributional assumptions beyond slow variations and can extend to other microgrid management tasks.
Abstract
from arXiv · showhide
Distribution microgrids are being challenged by reverse power flows and voltage fluctuations due to renewable generation, demand response, and electric vehicles. Advances in photovoltaic (PV) inverters offer new opportunities for reactive power management provided PV owners have the right investment incentives. In this context, reactive power compensation is considered here as an ancillary service. Accounting for the increasing time-variability of distributed generation and demand, a stochastic reactive power compensation scheme is developed. Given uncertain active power injections, an online reactive control scheme is devised. This scheme is distribution-free and relies solely on power injection data. Reactive injections are updated using the Lagrange multipliers of a second-order cone program. Numerical tests on an industrial 47-bus microgrid and the residential IEEE 123-bus feeder corroborate the reactive power management efficiency of the novel stochastic scheme over its deterministic alternative, as well as its capability to track variations in solar generation and household demand.
I. INTRODUCTION
Reactive power management in renewable-rich microgrids must address uncertain injections, reverse flows, and voltage fluctuations. The paper develops a stochastic, online framework that compensates controllable reactive power through an ancillary-service market and a convergent optimization scheme.
- Motivation: Renewable generation, elastic loads, and electric-vehicle charging create uncertain active-power variations, reverse flows, over-voltages, and voltage sags in distribution microgrids.PV output can vary by 15% of nameplate capacity within one-minute intervals.
- Motivation: Conventional voltage-regulation devices are limited by operational costs, discrete actions, and slow response, with islanded microgrids also lacking centralized fast-reacting generators.The paper motivates subsidized reactive-power control by distributed-generation units as an alternative.
- Research gap: Existing reactive-control schemes rely on approximate grid models and assume active-power injections are precisely known and unchanged during the control period.These assumptions are less realistic for microgrids with high renewable penetration.
- Contributions: The proposed framework models noisy and delayed load and renewable-generation estimates as stochastic processes and reimburses PV owners for reactive-power support.Reactive injections are selected as minimizers of expected reactive-power compensation cost.
- Contributions: A provably convergent stochastic-approximation algorithm computes subgradients through the dual SOCP and updates reactive PV injections using a thresholding rule.Numerical tests use industrial and residential microgrids with real solar-generation and demand data.
II. SYSTEM MODEL
The paper models the microgrid as a radial tree using a branch-flow representation derived from the full AC model. Nodal injections, line flows, voltages, and currents are constrained to support reactive-power control within prescribed voltage limits.
- Network representation: The microgrid has N + 1 buses and is represented as a radial tree rooted at substation bus 0, with each non-root bus connected to a unique parent.Each directed edge feeding a non-root bus is indexed by that bus.
- Network representation: For each non-root bus, the model defines squared voltage magnitude, complex power injection, line impedance, squared current magnitude, and sending-end line power flow.The branch-flow variables describe both nodal quantities and line quantities.
- Branch-flow model: The branch-flow equations follow from power conservation, squared Kirchhoff voltage relations, and current computations after eliminating voltage and current phases from the full AC model.The child set Cn collects buses whose parent is bus n.
- Nodal injections: Active and reactive nodal injections are decomposed into generation and consumption components, allowing distributed-generation buses to generate active power and provide controllable reactive support.Pure-load and shunt-capacitor buses are represented as special cases.
- Operational constraints: Bus voltage magnitudes are constrained to a prespecified range, typically ±5% of nominal voltage, and the resulting branch-flow model supports the stochastic reactive-control formulation.The paper introduces the control scheme after establishing the model and voltage constraints.
III. PROBLEM FORMULATION
The formulation treats reactive power compensation as an ancillary service for minimizing expected losses and maintaining voltage limits under uncertain, time-varying injections. PV inverter capabilities and compensation prices define the feasible control region and market objective.
- Objective: Reactive power management chooses reactive injections to minimize distribution-line losses while keeping bus voltages within prescribed limits.The stated example voltage range is ±5% of nominal values.
- Uncertainty: The controller operates over short intervals while active and reactive injections are observed as stochastic, noise-contaminated quantities.The framework models injections as time-independent draws from probability distributions and uses observations rather than exact forecasts.
- Ancillary service: Reactive compensation is formulated as an ancillary market in which PV owners are reimbursed for providing reactive-power support.The objective trades power losses against compensation payments using normalized support prices.
- Feasible region: PV inverter oversizing creates a time-invariant reactive injection region, allowing support independently of instantaneous PV output.Without oversizing, the feasible region varies with generation and may provide no reactive power at maximum solar output.
- Solution motivation: The expected-cost problem is nontrivial even with a known joint injection distribution, motivating stochastic approximation for practical solution.The stochastic formulation is intended to produce smoother control actions than instantaneous decisions.
IV. STOCHASTIC APPROXIMATION SOLVER
The solver uses online stochastic approximation to update reactive-injection estimates as new injection observations arrive. Composite-objective mirror descent replaces the expected problem with tractable local updates using subgradients and step sizes.
- Online updates: Successive reactive-injection estimates are updated whenever a new active and reactive injection datum becomes available.The estimates target the minimizer of the expected compensation objective.
- Optimization step: Composite-objective mirror descent forms each iterate by minimizing a convex local problem around the previous estimate.The update uses a subgradient evaluated at the previous iterate and an appropriately selected positive step size.
- Cost approximation: The stochastic approximation method substitutes an instantaneous cost for the original expected cost before constructing the online surrogate.The per-time minimizer is the optimal reactive injection for that realization, whereas online optimization minimizes a locally tight upper bound.
- Implementation: Implementing the control scheme requires solving the local minimization and computing the cost subgradient.These are identified as the two practical implementation issues.
A. Closed-Form Minimizer for (13)
The local minimization decouples across reactive-injection entries and yields a closed-form thresholding rule. The rule suppresses small updates, shifts intermediate values, and saturates large injections.
- Separable minimization: Completing the square converts the local optimization into a separable minimization over the entries of reactive injection.Each scalar problem depends on the corresponding entry of the transformed iterate.
- Closed form: The scalar minimizer is available in closed form through the KKT conditions.This result is stated as Proposition 1 for the univariate minimization.
- Thresholding: When |y_n,t| is below η_t c_n, the rule sets reactive injection at bus n to zero.The compensation price and step size determine the no-injection threshold.
- Thresholding: For larger |y_n,t|, the update either saturates or equals y_n,t − sign(y_n,t)η_t c_n in the intermediate regime.Thus, once a valid subgradient is available, the reactive-injection update is simple to evaluate.
B. Efficient Subgradient Computation
The subgradient is obtained from the dual of an SOCP representation of instantaneous power loss. Under exactness and regularity assumptions, dual multipliers support the online update, and the resulting scheme is data-driven and extensible.
- SOCP representation: Instantaneous loss is represented as the optimum value of an SOCP whose inputs are active injections and reactive injections.The loss function is convex in reactive injection because it is a perturbation function.
- Assumptions: The formulation assumes an exact convex relaxation and a strictly feasible point, and these assumptions were verified numerically in the reported instances.The paper states that the assumptions are not analytically supported in the presented development.
- SOCP representation: Eliminating selected primal variables expresses power-flow quantities affinely in auxiliary variables, while voltage and current constraints become conic constraints.This produces an equivalent SOCP formulation for computing the loss value.
- Dual subgradient: The negative optimal dual variables associated with the reactive-injection constraints provide a subgradient of the loss function.Strong duality and sensitivity analysis justify extracting the subgradient from the dual SOCP.
- Dual subgradient: The control update sets the subgradient using the optimal dual multiplier, after which the closed-form thresholding rule can be applied.Complementary slackness also supplies an exactness certificate through the dual variables associated with the conic constraints.
- Scope: The resulting scheme uses real-time operation data without distributional assumptions and can track slowly time-varying injection statistics.The framework is also identified as applicable to voltage-deviation and conservation-voltage-regulation tasks.
C. Algorithm Convergence
The stochastic control algorithm converges toward the optimum stochastic power loss, with high-probability bounds and sublinear regret. Its guarantees depend on the chosen step size and parameters governing the reactive-injection region and subgradient norms.
- C. Algorithm Convergence: The iterates converge toward a minimizer of the stochastic optimization problem under the stated proposition.Proposition 2 characterizes convergence for the iterates defined by the algorithm.
- C. Algorithm Convergence: The expected power loss of the averaged reactive-control iterate converges to the optimum stochastic power loss at rate O(1/T).This is the mean-value convergence guarantee identified after Proposition 2.
- C. Algorithm Convergence: The averaged cost remains close to the optimum with high probability, and the algorithm therefore has sublinear regret.The high-probability statement is associated with the bound in Proposition 2.
- C. Algorithm Convergence: A constant step size is permitted when the time horizon T is known in advance, while a time-decaying step size works when T is unknown.The unknown-horizon option incurs a slight degradation in performance.
- C. Algorithm Convergence: The convergence bounds depend on D and L, with D linked to PV reactive-power capabilities and L bounded when the feasible injection region is compact.Precisely knowing L may be practically unrealistic.
V. NUMERICAL TESTS
Numerical tests compare stochastic and deterministic reactive power control on industrial and residential feeders under uncertain, delayed, and real solar and demand data. The stochastic scheme achieves lower costs and tracks changing generation and load conditions.
- 47-bus industrial feeder: The 47-bus industrial feeder experiment compares deterministic control with stochastic control using observed active injections and evaluates cost on the true system state.The feeder is from South California Edison; all SOCP relaxations were feasible and exact.
- 47-bus industrial feeder: After 20 iterations, the stochastic algorithm converges to low cost, while the deterministic alternative remains at consistently higher costs in one realization.Each reactive control run completed within 1.2 seconds on the stated MATLAB/CVX setup.
- 47-bus industrial feeder: 28.7, 39.7, 41.8, 44.9, and 45.6 $/h are the stochastic savings for η = 1, 2, 2.5, 3.5, and 4, respectively, averaged over 40 realizations.Larger step sizes slow convergence but reduce steady-state cost, trading off transient behavior and statistical tracking.
- Real solar generation: With real solar generation and delayed observations, the stochastic scheme approaches ideal control after convergence and tracks solar variations, yielding a 12.7$/h benefit over deterministic control.The ideal controller uses the instantaneous state and serves as a practically infeasible lower bound.
- Real solar generation: During daylight, the stochastic scheme tracks the steady solar ramp from 7.30–9.15am and cloud-related variations while remaining slightly above ideal-control cost.The experiment uses 30-second control intervals and a 30-second observation delay.
- Real solar and demand data: On the IEEE 123-bus residential feeder, real solar and household demand data show that the stochastic scheme successfully tracks both solar and load variations.The experiment uses a one-minute control period and one-minute delayed observations.
VI. CONCLUDING REMARKS
The paper develops a stochastic reactive power compensation scheme that updates PV-inverter injections in real time under uncertain microgrid conditions. Tests show convergence within 10–20 iterations, lower cost than a myopic deterministic alternative, and successful tracking of system variations.
- VI. CONCLUDING REMARKS: The proposed stochastic scheme updates reactive power injections from PV inverters in real time.It is designed for uncertain microgrid states and builds on convex relaxation and online convex optimization.
- VI. CONCLUDING REMARKS: 10–20 iterations were sufficient for the novel control scheme to converge in tests on practical microgrids.
- VI. CONCLUDING REMARKS: The achieved reactive power management cost was consistently lower than that of the myopic deterministic alternative.
- VI. CONCLUDING REMARKS: Experiments with real solar-generation and load-consumption data showed successful tracking of underlying system variations and approach toward ideal reactive control.
- VI. CONCLUDING REMARKS: The framework imposes no distributional assumptions on active injections apart from requiring slow variations and adjusts dynamically to microgrid operation data.
APPENDIX
The appendix proves a proposition by applying KKT conditions and subgradient properties to characterize the relevant cases for the canonical problem. In the zero solution case, feasibility occurs only when a lies within [−b, b].
- APPENDIX: The proof applies KKT conditions involving optimal Lagrange multipliers for box constraints and a subgradient of |x|.
- APPENDIX: For x ≠ 0, the subgradient of |x| equals sign(x), while at zero its magnitude is at most 1.
- APPENDIX: The proof separates cases according to whether the optimizer ˆx is positive, negative, or zero.
- APPENDIX: When ˆx = 0, the conditions imply ξ = 0 and a = bs(ˆx).
- APPENDIX: The zero case occurs only when a ∈[−b, b], because the subgradient at zero satisfies |s(0)| ≤ 1.