Transmission Volt/VAR Optimization Framework

• Transmission Volt/VAR Optimization is a coordinated framework that integrates AC power-flow physics and discrete device models to manage voltage levels and reactive power.
• It employs methodologies such as continuous relaxation, discrete rounding, and real-time control to minimize voltage deviations and operational losses.
• Advanced strategies including hierarchical control and deep reinforcement learning enable effective DER integration and enhance grid reliability.

Transmission Volt/VAR Optimization (VVO) frameworks coordinate voltage levels and reactive power (VAR) flows throughout transmission networks to ensure voltage stability, minimize losses, and optimize resource utilization. These frameworks must account for device discreteness, AC power-flow physics, operational limits, and increasingly, the integration of high-penetration distributed energy resources (DERs). Recent advances include optimal planning tools, hierarchical real-time control architectures, relax–round–resolve heuristics, and deep reinforcement learning approaches.

1. Formal Problem Definition and Device Modeling

Transmission VVO frameworks are typically formulated as mixed-integer nonlinear programming (MINLP) problems over a network graph $G = (N, E)$ of buses ($N$) and branches ($E$). The principal decision variables encompass:

• Continuous variables: bus voltages $V_i = v_i e^{j\theta_i}$, generator setpoints $p_i$, $q_i$, and power flows $S_{ij}$.
• Discrete variables: shunt susceptance settings $b_i \in B_i$ (capacitor/inductor modules) and on-load tap changer (OLTC) transformer tap ratios $\tau_{ij} \in \mathcal T_{ij}$, where $B_i$ and $\mathcal T_{ij}$ are device-specific discrete sets.

Key constraints encode AC power-flow (Kirchhoff’s equations), generator and branch limits, thermal constraints, voltage and angle limits, and manufacturer-accurate device characteristics. The objective function comprises weighted terms penalizing voltage deviations, VAR usage, active-power redispatch, and (optionally) generation costs: $$\psi(v, q, p) = \lambda_v \sum_{i\in N}(v_i-\hat v_i)^2 + \lambda_q \sum_{i\in N}(q_i-\hat q_i)^2 + \lambda_p \sum_{i\in N}(p_i-\hat p_i)^2 + \lambda_c \sum_{i\in N} c_i(p_i).$$ (Tong et al., 29 Jan 2026, Nguyen et al., 2019)

Device models must accurately represent actual step sizes and initial states, for example OLTC tap sets such as $\mathcal{T}_{ij} = \{0.9,\,0.90625,\dots,1.1\}$ covering $\pm16$ steps of 0.625%, and shunt capacitors banked in modules (e.g., $B_i = \{0, 0.1, 0.2, 0.3\}$ in per unit on a 100 MVA base) (Tong et al., 29 Jan 2026).

2. Algorithmic Frameworks: Relaxation, Rounding, and Real-Time Control

Direct solution of large-scale VVO MINLPs is computationally prohibitive. Current frameworks implement a multi-phase pipeline:

1. Continuous Relaxation: Discrete variables $b_i$, $\tau_{ij}$ are relaxed to continuous intervals. The resulting nonconvex AC-OPF is solved (e.g., via Ipopt), yielding fractional device settings.
2. Discrete Rounding: Relaxed outputs $\tilde{b}_i$, $\tilde{\tau}_{ij}$ are discretized by nearest-element projection onto $B_i$, $\mathcal{T}_{ij}$.
3. Feasibility Restoration: The AC-OPF is resolved with fixed discrete device choices, ensuring full physico-operational feasibility (“resolve” phase). (Tong et al., 29 Jan 2026)

For high-DER sub-transmission planning, scenario-based approaches leverage representative “worst violation” power-flow snapshots, clustering and principal component analysis to minimize the investment in new VAR devices without compromising voltage security (Nguyen et al., 2019).

3. Hierarchical and Coordinated Multi-Level Control Architectures

Modern VVO frameworks must co-optimize transmission and distribution-level resources. The Coordinative Real-time Sub-Transmission Volt-Var Control Tool (CReST-VCT) establishes a two-level hierarchy:

• Stage 1 (EMS/CReST-VCT): Determines sub-transmission VAR resources and aggregated feeder-level reactive support, enforcing sub-transmission constraints and scheduling VAR exchanges with distribution roots.
• Stage 2 (DMS/VLSM): Disaggregates aggregate VAR demands to individual DERs (e.g., PV inverters, storage), optimizing feeder voltage and resource costs.

Data exchanges between EMS and DMS follow standardized protocols (e.g., DNP3-SA over WAN, IEC 61850 within substations), with latency requirements regulated to strict operational timelines (e.g., sub-300s real-time cycle). Cybersecurity measures, including TLS, certificate-based mutual authentication, and role-based access, are implemented throughout (Nguyen et al., 2021).

This architecture enables co-optimization of DER capabilities, resulting in loss reductions (8–12%), high DER reactive utilization (>85%), and near-elimination of voltage violations under high renewable penetration (Nguyen et al., 2021).

4. Reinforcement Learning and Scalable Computational Strategies

Deep reinforcement learning (DRL) introduces data-driven control to VVO by framing the problem as a Markov decision process (MDP) with state vector $s_t$ (voltages, demands), action vector $a_t$ (DER setpoints, device switching), transition dynamics given by AC power-flow, and reward penalizing voltage violations: $r_t = -V_{\mathrm{vio}}(t).$ The RLlib-IMPALA framework utilizes distributed actor-learner architecture (V-trace off-policy corrections) enabling parallelized trajectory sampling and policy updates across multiple compute cores (Selim et al., 2024). This reduces wall-clock training for zero-violation policies from hours to minutes and outperforms standard DRL baselines (e.g., PPO, SAC) in reward and scalability.

The integration with industry-standard simulation backends (e.g., OpenDSS) and Ray distributed computing allows handling of real-time operational problems on large networks, though core-contention and ACOPF solvability remain limiting for networks above 5,000 nodes without further parallel scaling (Selim et al., 2024).

5. Performance Metrics and Validation Methodologies

Framework performance is evaluated using voltage profile deviation (MAE from 1 p.u.), mean absolute reactive generation ($\mathrm{MAE}_q$), active redispatch $\Delta_p$, cost impact $\%\Delta c$, loss reduction, DER utilization, and cycle-time including computational and communication latency (Tong et al., 29 Jan 2026, Nguyen et al., 2021, Nguyen et al., 2019).

Table: Selected Performance Outcomes

Metric	Result (Representative System)	Source
Voltage deviation MAE	10–50% reduction	(Tong et al., 29 Jan 2026)
Reactive generation MAE	Halved in best cases	(Tong et al., 29 Jan 2026)
Active power redispatch ($\Delta_p$)	<3% (modest)	(Tong et al., 29 Jan 2026)
Generation-cost savings ($\% \Delta c$)	0.1–3%	(Tong et al., 29 Jan 2026)
DER utilization	85%	(Nguyen et al., 2021)
Voltage violations eliminated	70 → <5 per day	(Nguyen et al., 2021)
RLlib-IMPALA speedup	×10 vs. SAC, ×2 vs. PPO	(Selim et al., 2024)

6. Practical Deployment, Robustness, and Future Extensions

Field-deployable VVO frameworks rely on:

• Realistic device models: Manufacturer step sizes and initial tap positions are essential for operator acceptance.
• Moderate device-movement limits: Restricting OLTC tap movements and capacitor banks expedites convergence with little loss of benefit.
• Warm starts: Using prior AC-OPF solutions to initialize optimization accelerates feasibility.
• Hierarchy and modularity: Decoupling transmission and distribution, encapsulating DER aggregation, and leveraging open-source interoperability (CIM, GridAPPS-D, OpenFMB).
• Scalability: Medium-sized networks (<3,000 buses) are tractable in seconds to minutes; further decomposition or distributed computation is required for larger grids (Tong et al., 29 Jan 2026, Nguyen et al., 2021).

Limitations include the absence of explicit N-1 security, dynamic/timed coordination between fast (inverter) and slow (OLTC/shunt) devices, and unmodeled FACTS devices and transient phenomena. Ongoing research extends these frameworks to security-constrained optimization and coordinated storage dispatch (Nguyen et al., 2019).

Indications from RL-based VVO frameworks suggest promising paths in hierarchical/distributed learning agents, transfer learning across topologies, and the explicit integration of forecast and uncertainty (Selim et al., 2024).

7. Comparative Analysis of Frameworks and Strategic Implications

Framework	Device Scope	Solution Methodology	Scale (Buses)	Key Strengths	Paper
MINLP with relax–round–resolve	OLTCs, shunt banks	Nonconvex ACOPF + discrete rounding	Up to 14,000	Realistic device modeling, AC feasibility	(Tong et al., 29 Jan 2026)
Scenario-based planning with clustering	All VAR assets, PV inverters	Scenario selection, MINLP, iterative rounding	125+	Investment cost minimization under high PV	(Nguyen et al., 2019)
RLlib-IMPALA DRL	Shunt, tap, DERs	Distributed RL, AC power-flow env	500–1,000+	Scalability, real-time control, zero-violation	(Selim et al., 2024)
CReST-VCT hierarchical	Transmission, distribution, DER	Two-level convex/nonconvex OPF	3,000+	EMS/DMS coordination, real-time deployment	(Nguyen et al., 2021)

Significant reductions in investment, violation count, and operational losses are reported across frameworks. Including VAR support from PV inverters or other DERs consistently decreases the need for new VAR device investments by 30–40% for the same voltage performance (Nguyen et al., 2019). A plausible implication is that integration of advanced inverter functionality is a critical enabler for cost-efficient VVO under high-renewable scenarios.

• "Volt/VAR Optimization in Transmission Networks with Discrete-Control Devices" (Tong et al., 29 Jan 2026)
• "Optimal Future Sub-Transmission Volt-Var Planning Tool to Enable High PV Penetration" (Nguyen et al., 2019)
• "Scalable Volt-VAR Optimization using RLlib-IMPALA Framework: A Reinforcement Learning Approach" (Selim et al., 2024)
• "EMS and DMS Integration of the Coordinative Real-time Sub-Transmission Volt-Var Control Tool under High DER Penetration" (Nguyen et al., 2021)