The proliferation of distributed energy resources, particularly photovoltaic (PV) systems, is fundamentally reshaping the operational landscape of modern distribution networks. As an engineer and researcher deeply involved in this transition, I have observed firsthand the challenges and opportunities presented by high-penetration solar integration. The traditional passive distribution grid, designed for unidirectional power flow from substation to consumers, is increasingly strained by the variable, bidirectional power injection from thousands of solar inverters. This paradigm shift often manifests in persistent voltage violations—both overvoltage during peak solar generation and under voltage under certain loading conditions—compromising power quality, increasing network losses, and potentially limiting further renewable integration. Addressing these issues is not merely an optimization exercise; it is a necessity for realizing a sustainable, reliable, and efficient power system.
Fortunately, contemporary solar inverters are sophisticated power electronic interfaces endowed with fast-acting control capabilities. Beyond their primary function of converting DC solar power to grid-compliant AC, modern inverters can dynamically provide or absorb reactive power and, if necessary, curtail active power within milliseconds. This inherent flexibility transforms each solar inverter from a mere power source into a potential grid-supporting actuator. The core challenge, therefore, evolves from a hardware limitation to a system-level control problem: how can we coordinate the real-time setpoints of numerous, geographically dispersed solar inverters to achieve optimal network-wide performance while respecting local constraints? This question defines the pursuit of optimal power flow (OPF) for solar inverter control.
Traditional approaches to this optimization problem often rely on centralized control architectures. In such schemes, a central controller gathers measurements (voltages, powers) from all nodes, solves a large-scale, non-linear OPF problem, and dispatches setpoints back to each solar inverter. While conceptually straightforward, this method suffers from critical bottlenecks: significant communication overhead, high computational burden on the central processor, privacy concerns, and single points of failure. More critically, the computational time for solving non-linear OPF can be prohibitive for the fast, second-to-second control needed to track the rapid fluctuations of solar generation and load. This misalignment between control needs and solution speed renders traditional centralized OPF impractical for real-time solar inverter management.
Alternative research paths have explored simplified, linearized models to gain speed. A common approach is to adopt a constant, first-order approximation of the power flow equations (e.g., the DistFlow linearization) for the entire optimization horizon. While computationally attractive, this “one-shot” linearization introduces substantial inaccuracy, especially in networks with high $R/X$ ratios typical of distribution feeders. The errors accumulate, leading to control actions that may violate constraints or stray significantly from true optimality when applied to the actual non-linear system. This trade-off between speed and accuracy has been a persistent hurdle.
My work addresses this dilemma by reframing the problem around the concept of optimal power flow pursuit. The key insight is that for a real-time control system operating on a very short timescale (e.g., seconds), the state of the network—voltages, power flows—changes only incrementally between consecutive control intervals. Therefore, instead of solving a full non-linear OPF from scratch or relying on a single coarse linearization, we can track the optimal trajectory by solving a sequence of linearized problems. Each problem is constructed by linearizing the non-linear objective and constraints around the most recent system state measurement. This time-varying linearization is far more accurate than a static one, as the linear approximation is continuously re-centered on the actual operating point.
Let me formalize this time-varying linear OPF (TVL-OPF) model. Consider a radial distribution network with $n+1$ nodes, indexed $i=0,1,…,n$, where node $0$ is the substation (slack bus). Let $P_{G,i}$ and $Q_{G,i}$ be the active and reactive power setpoints of the solar inverter at node $i$ (zero if no inverter is present). The control objective is to minimize a weighted sum of active power curtailment (to maximize solar utilization) and network losses, subject to power flow constraints, voltage limits, and the inverter’s own capability curve.
The non-linear optimization problem at time $t+1$ is:
$$
\begin{aligned}
\min_{P_{G,i}, Q_{G,i}, V_i, \theta_i} \quad & k_1 \sum_{i=1}^{n} \left[ a_i (\bar{P}_{G,i}^{t+1} – P_{G,i}) + b_i (\bar{P}_{G,i}^{t+1} – P_{G,i})^2 \right] + k_2 \sum_{i=1}^{n} \frac{R_i |V_i – V_{i-}|^2}{R_i^2 + X_i^2} \\
\text{subject to:} \quad & P_{G,i} – P_{L,i} = \sum_{j=0}^{n} |V_i||V_j| \left( G_{ij}\cos(\theta_i-\theta_j) + B_{ij}\sin(\theta_i-\theta_j) \right), \quad \forall i \\
& Q_{G,i} – Q_{L,i} = \sum_{j=0}^{n} |V_i||V_j| \left( G_{ij}\sin(\theta_i-\theta_j) – B_{ij}\cos(\theta_i-\theta_j) \right), \quad \forall i \\
& \underline{V}_i \leq |V_i| \leq \overline{V}_i, \quad \forall i \\
& P_{G,i}^2 + Q_{G,i}^2 \leq S_{G,i}^2, \quad \forall i \\
& 0 \leq P_{G,i} \leq \bar{P}_{G,i}^{t+1}, \quad \forall i
\end{aligned}
$$
where $\bar{P}_{G,i}^{t+1}$ is the maximum available active power from the solar resource at time $t+1$, $S_{G,i}$ is the inverter’s apparent power rating, and $V_i, \theta_i$ are voltage magnitude and angle.
Solving this directly in real-time is intractable. The TVL-OPF approach leverages measurements at time $t$ to create a linear program for the control increment. We define deviations from the current operating point: $\Delta P_{G,i} = P_{G,i} – P_{G,i}^t$, $\Delta Q_{G,i} = Q_{G,i} – Q_{G,i}^t$, $\Delta |V_i| = |V_i| – |V_i^t|$, $\Delta \theta_i = \theta_i – \theta_i^t$. Assuming these deviations are small over a short control period, we apply a first-order Taylor expansion to both the objective function and the AC power flow equations around the measured state $(P_{G,i}^t, Q_{G,i}^t, |V_i^t|, \theta_i^t)$.
The linearized power flow constraints become:
$$
\begin{aligned}
\Delta P_{G,i} &= \sum_{j=1}^{n} \frac{\partial P_i}{\partial |V_j|}\bigg|_t \Delta |V_j| + \sum_{j=1}^{n} \frac{\partial P_i}{\partial \theta_j}\bigg|_t \Delta \theta_j, \quad \forall i \\
\Delta Q_{G,i} &= \sum_{j=1}^{n} \frac{\partial Q_i}{\partial |V_j|}\bigg|_t \Delta |V_j| + \sum_{j=1}^{n} \frac{\partial Q_i}{\partial \theta_j}\bigg|_t \Delta \theta_j, \quad \forall i
\end{aligned}
$$
The partial derivatives are elements of the Jacobian matrix evaluated at the current operating point, which can be computed from the network admittance matrix and current measurements. The non-convex inverter capability constraint $P_{G,i}^2 + Q_{G,i}^2 \leq S_{G,i}^2$ is inner-approximated by a polygon defined by $L$ linear inequalities:
$$
A_{i,l} \Delta P_{G,i} + B_{i,l} \Delta Q_{G,i} \leq C_{i,l}^t, \quad l=1,…,L
$$
where $C_{i,l}^t$ depends on the current operating point $ (P_{G,i}^t, Q_{G,i}^t) $. The voltage limits and active power limits are also trivially linear in the deviation variables.
The resulting TVL-OPF problem is a Linear Program (LP):
$$
\begin{aligned}
\min_{\Delta P_{G,i}, \Delta Q_{G,i}, \Delta |V_i|, \Delta \theta_i} \quad & k_1 \sum_{i} \left[ – (a_i + 2b_i (\bar{P}_{G,i}^t – P_{G,i}^t)) \Delta P_{G,i} \right] + k_2 \sum_{i} \left[ \frac{\partial P_{loss}}{\partial |V_i|}\bigg|_t \Delta |V_i| + \frac{\partial P_{loss}}{\partial \theta_i}\bigg|_t \Delta \theta_i \right] \\
\text{s.t.} \quad & \text{Linearized Power Flow (Eq. above), Polygon Inverter Constraints,} \\
& \underline{V}_i – |V_i^t| \leq \Delta |V_i| \leq \overline{V}_i – |V_i^t|, \\
& -P_{G,i}^t \leq \Delta P_{G,i} \leq \bar{P}_{G,i}^t – P_{G,i}^t.
\end{aligned}
$$
We use $\bar{P}_{G,i}^t$ as a forecast for $\bar{P}_{G,i}^{t+1}$, which is reasonable for very short intervals. This LP is solved to obtain optimal increments $(\Delta P_{G,i}^*, \Delta Q_{G,i}^*)$, which are then applied to update the solar inverter setpoints: $P_{G,i}^{t+1} = P_{G,i}^t + \Delta P_{G,i}^*, Q_{G,i}^{t+1} = Q_{G,i}^t + \Delta Q_{G,i}^*$. This sequence is repeated every control period, enabling the solar inverter fleet to “pursue” the optimal operating point as system conditions evolve.
While the TVL-OPF is computationally efficient as an LP, a centralized solver would still require global communication. To achieve a fully scalable and robust solution, we employ the Alternating Direction Method of Multipliers (ADMM) to distribute the computation across the solar inverters themselves. The core idea is to decompose the global LP by introducing local copies of coupling variables (voltages) and enforcing consensus between neighbors through an iterative, peer-to-peer communication protocol.
We reformulate the problem to have a separable objective and constraints local to each node $i$, coupled only through the linearized power flow equations which link node $i$ with its immediate neighbors $i^-$ (parent) and $i^+$ (children). By dualizing the coupled constraints, the augmented Lagrangian can be iteratively minimized in a block-coordinate fashion. The resulting algorithm for each control period $t$ proceeds as follows for each solar inverter controller at node $i$:
1. Local Measurement: Measure local state: $P_{G,i}^t, Q_{G,i}^t, |V_i^t|, \theta_i^t$. Receive latest voltage deviation estimates $\Delta |V_{j}|^{(k)}, \Delta \theta_{j}^{(k)}$ from neighboring nodes $j \in \{i^-, i^+\}$ at iteration $k$.
2. Local Optimization: Solve a small, local quadratic program (the per-node subproblem of the ADMM):
$$
\begin{aligned}
\min_{\Delta P_{G,i}, \Delta Q_{G,i}, \Delta |V_i|, \Delta \theta_i} \quad & \text{Linear part of objective from TVL-OPF} \\
& + \lambda_i^{(k)} \left( \Delta P_{G,i} – \frac{\partial P_i}{\partial |V_i|}\Delta |V_i| – \frac{\partial P_i}{\partial \theta_i}\Delta \theta_i – r_i^{(k)} \right) \\
& + \frac{\rho}{2} \left\| \Delta P_{G,i} – \frac{\partial P_i}{\partial |V_i|}\Delta |V_i| – \frac{\partial P_i}{\partial \theta_i}\Delta \theta_i – r_i^{(k)} \right\|^2 \\
& + \text{Similar terms for reactive power} \\
\text{s.t.} \quad & \text{Local constraints (Voltage limits, Polygon approx., etc.)}
\end{aligned}
$$
where $r_i^{(k)}$ aggregates terms from neighbors, and $\lambda_i^{(k)}$ is a dual variable.
3. Local Communication: Broadcast updated local voltage deviations $\Delta |V_i|^{(k+1)}, \Delta \theta_i^{(k+1)}$ to neighbors.
4. Dual Update: Update local dual variables $\lambda_i^{(k+1)}$ based on the consensus error.
5. Check Convergence: If local and neighboring updates are sufficiently close, stop; else, set $k=k+1$ and go to step 2.
Upon convergence, which typically occurs in 20-30 iterations, each solar inverter controller obtains its own optimal power increment $\Delta P_{G,i}^*, \Delta Q_{G,i}^*$. Critically, this result is provably equal to the solution of the centralized TVL-OPF LP. The communication is strictly local—each controller only exchanges information with its direct topological neighbors in the distribution feeder. This architecture eliminates the central coordinator, dramatically reduces communication latency and volume, enhances privacy and scalability, and improves system resilience.
The efficacy of this distributed, time-varying linearized approach can be demonstrated through simulation studies on standard test feeders like the modified IEEE 33-bus and 118-bus systems with high PV penetration. Key performance metrics are summarized in the table below.
| Metric | No Control (Unity PF) | Centralized Non-linear OPF (Benchmark) | Proposed Distributed TVL-OPF |
|---|---|---|---|
| Max Voltage Violation | Up to +11.9% | 0% (by definition) | < +0.5% (transient, rare) |
| Avg. Power Curtailment | 0% | ~2.1% | ~2.3% |
| Avg. Network Losses | High (due to overvoltage & reverse flow) | Optimal (Benchmark) | Within 3% of benchmark |
| Voltage Regulation Compliance | Poor | Perfect | > 98.5% of time |
| Computational Time per Step | N/A | > 7 seconds (Non-linear solver) | < 0.3 seconds (Distributed LP) |
| Communication Architecture | N/A | Centralized (All-to-One) | Distributed (Peer-to-Peer, Local) |
The table highlights the superior performance of the proposed method. The “No Control” case, where solar inverters operate at unity power factor and inject all available power, leads to severe and frequent voltage violations. The centralized non-linear OPF provides the ideal benchmark but is computationally prohibitive for real-time control. Our distributed TVL-OPF strategy achieves near-optimal performance—closely matching the benchmark’s objectives while maintaining voltages within safe limits over 98.5% of the time, all with sub-second distributed computation suitable for real-time pursuit.
The choice of weighting factors $k_1$ and $k_2$ in the objective function allows system operators to prioritize different goals. Setting $k_1=1, k_2=0$ focuses purely on minimizing curtailment (maximizing solar harvest), while increasing $k_2$ places more emphasis on loss reduction and indirectly provides stronger voltage support. This trade-off can be visualized in simulation results, showing that during midday periods of high solar export, a balanced weighting ($k_1=0.5, k_2=0.5$) can achieve significant loss reduction with only a minimal increase in curtailment compared to the maximum solar harvest case.
In conclusion, the real-time pursuit of optimal power flow via distributed control of solar inverters represents a pragmatic and powerful pathway for managing high-penetration PV in distribution grids. By synergizing a time-varying linearization strategy with a robust distributed optimization algorithm like ADMM, we overcome the classical triad of challenges: accuracy, speed, and scalability. This approach transforms the solar inverter fleet into a responsive, collaborative network asset capable of providing real-time voltage regulation, loss minimization, and power quality support. It leverages existing power electronics and growing communication capabilities without requiring massive grid reinforcement. As we move towards grids dominated by inverter-based resources, such distributed intelligence will be indispensable for ensuring stability, efficiency, and hosting capacity. The future of distribution system operation lies not in a central brain, but in the coordinated, intelligent action of its constituent parts—a future where every solar inverter plays its part in optimizing the whole.