DC Optimal Power Flow

• DC Optimal Power Flow is a linear/quadratic optimization model that dispatches generation to meet demand at minimum cost while incorporating linearized network constraints.
• It employs a DC approximation with fixed voltage magnitudes and small angle differences to convert nonlinear power flows into a tractable linear formulation.
• Recent research integrates stochastic extensions, GPU acceleration, and learning-based methods to enhance scalability and address the AC-feasibility gap.

DC Optimal Power Flow (DC-OPF) is the linear or quadratic optimization model that dispatches active power generation to meet demand at minimum cost while enforcing generator bounds, linearized network power-balance equations, and transmission flow limits under the DC power-flow approximation. It is a key operational tool for power system operators, is widely used for market clearing and fast operational planning, and is embedded as a subproblem in more challenging optimization problems such as line switching, stochastic dispatch, and security-constrained formulations (Rafiei et al., 2024, Baker, 2019).

1. Canonical problem statement and DC approximation

The standard DC-OPF model assumes voltage magnitudes are fixed and close to $1$ p.u., angle differences are small, line resistances and losses are negligible relative to reactances, and reactive power is ignored. Under these assumptions, active power flows become linear functions of bus phase-angle differences, and the optimization is posed over generator active powers and bus angles, or equivalently over generator powers alone in a PTDF representation (Baker, 2019, Pan et al., 2019).

A canonical angle-based formulation minimizes either linear cost $\sum_{g\in\mathcal{G}} c_g p_g$ or quadratic cost $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$ subject to

$B\theta + p - d = 0,$

$$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$

$p_g^{\min} \le p_g \le p_g^{\max},$

and a fixed reference angle $\theta_{\mathrm{ref}}=0$. In incidence–susceptance notation, the reduced bus susceptance Laplacian is $B = E^T \Omega E$, with $\Omega=\operatorname{diag}(b)$ and $b_\ell=1/x_\ell$ (Rafiei et al., 2024).

The same model can be expressed in PTDF form by eliminating angles. Then line flows are linear functions of net injections,

$\sum_{g\in\mathcal{G}} c_g p_g$0

or, with a slack-bus convention, $\sum_{g\in\mathcal{G}} c_g p_g$1. The optimization becomes a convex LP or QP with a global power-balance constraint, linear line-flow constraints, and box bounds on generators (Rafiei et al., 2024, Ng et al., 2018).

Several papers use closely related matrix forms. In one common notation, $\sum_{g\in\mathcal{G}} c_g p_g$2 and $\sum_{g\in\mathcal{G}} c_g p_g$3 with $\sum_{g\in\mathcal{G}} c_g p_g$4 (Mohammadi et al., 2014). In another, line flows are written as $\sum_{g\in\mathcal{G}} c_g p_g$5 under uncertain net load (Pena-Ordieres et al., 2019). These are not distinct models; they are algebraic variants of the same DC approximation.

2. Optimality structure, active sets, and prices

Because DC-OPF is an LP for linear costs and a convex QP for quadratic costs, its feasible set is polyhedral in the standard formulations, and strong duality typically holds under Slater-type conditions. This makes KKT systems, dual variables, and active-set structure central to both analysis and algorithms (Rafiei et al., 2024, Pena-Ordieres et al., 2019).

For a fixed uncertainty realization in the linear-cost case, the DC-OPF solution is a basic feasible solution in non-degenerate settings. If $\sum_{g\in\mathcal{G}} c_g p_g$6 denotes the $\sum_{g\in\mathcal{G}} c_g p_g$7 active inequality constraints, then the basis matrix

$\sum_{g\in\mathcal{G}} c_g p_g$8

defines an affine policy

$\sum_{g\in\mathcal{G}} c_g p_g$9

This yields the piecewise-affine structure exploited in statistical learning approaches: a basis policy is optimal exactly on the region of uncertainty realizations sharing the same active set (Ng et al., 2018).

The associated KKT system also provides an explicit pricing interpretation. With multipliers $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$0 and $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$1 on upper and lower line limits and scalar $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$2 for system balance, the locational marginal price at bus $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$3 is

$\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$4

If a generator is not at a bound, its marginal generation cost equals the LMP at its bus; binding generator limits introduce the usual upper- or lower-bound multipliers (Ng et al., 2018).

Dual reformulations of DC-OPF also play a major role in scalable algorithms. One line of work normalizes the primal variables to an $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$5 box, forms a Lagrangian dual, and solves the inner minimization in closed form using the dual norm identity

$\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$6

yielding an explicit concave dual objective for projected ascent (Rafiei et al., 2024). This suggests that much of DC-OPF’s algorithmic tractability arises not only from convexity, but from unusually explicit primal–dual structure.

3. Uncertainty, stochasticity, and security constraints

A large research thread studies DC-OPF under uncertainty in loads, renewable injections, and contingencies. One formulation introduces affine AGC recourse

$\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$7

and imposes a joint chance constraint requiring that all generator and line constraints be satisfied simultaneously with probability at least $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$8. The resulting problem is reformulated through a smooth sample-based quantile constraint and solved by an $\sum_{g\in\mathcal{G}}(a_g p_g^2+b_g p_g+c_g)$9-type trust-region method with convex QP subproblems and lazy constraint generation (Pena-Ordieres et al., 2019).

Distributionally robust variants replace a fixed probability law with an ambiguity set over scenario probabilities. A multi-stage formulation for uncertain renewable output uses

$B\theta + p - d = 0,$0

and optimizes against the worst-case expectation over $B\theta + p - d = 0,$1. The Bellman recursion is dualized into an SOCP, and the cost-to-go is approximated by convex quadratic surrogates fitted over interpolation points (Yang et al., 2023).

Scenario-based two-stage stochastic DC-OPF introduces first-stage dispatch and second-stage redispatch variables for many sampled load realizations. A central difficulty is that approximate second-stage policies must remain feasible for line limits, power balance, and generator bounds. One proposed remedy uses a gauge map on the scenario-specific polytope $B\theta + p - d = 0,$2, with

$B\theta + p - d = 0,$3

so that feedforward policies output only feasible recourse decisions (Zhang et al., 2023).

Security-constrained extensions enlarge the feasible set from one operating point to many contingency states. A preventive SCOPF formulation in a sparse device-node model enforces node balance and phase consistency in every contingency $B\theta + p - d = 0,$4, while requiring non-contingency device schedules to remain shared across contingencies. In the paper’s example, AC-line outages are represented by contingency-specific parameters $B\theta + p - d = 0,$5, so only the contingency-affected device type changes across scenarios (Degleris et al., 2024).

4. Computational methods and scalable solution architectures

Traditional CPU-based routines such as simplex are reported to have saturated in speed and to be hard to parallelize for repeated large-scale DC-OPF solves. A GPU-oriented alternative reformulates DC-OPF as a dual ascent problem over normalized box-constrained primal variables, solves the inner minimization analytically, and iterates the dual with projected gradient ascent. The dominant kernels are sparse matrix–vector products and elementwise operations, which map naturally to GPUs. On IEEE 2000-, 4601-, and 10000-bus systems, this method produced reliable and tight lower bounds and achieved, at best, a $B\theta + p - d = 0,$6 speedup over a conventional solver (Rafiei et al., 2024).

For contingency-constrained multi-period problems, a different GPU strategy uses proximal message passing in a sparse device-node model. The method is an ADMM variant with local device proximal operators, node-average and residual projections, and no outer linear-system solves. Implemented purely in PyTorch and run entirely on the GPU, it is end-to-end differentiable and reported well over $B\theta + p - d = 0,$7 speedups on large test cases; with $B\theta + p - d = 0,$8 up to $B\theta + p - d = 0,$9 contingencies it was reported to be $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$0 faster than Mosek, and it solved a case with over $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$1 million variables in under a minute on a single A100 GPU (Degleris et al., 2024).

Distributed algorithms attack DC-OPF from yet another direction. One fully distributed method iteratively updates bus-level angles, local multipliers, generator outputs, and line-limit multipliers using only neighbor communication, and proves convergence to the centralized optimum under suitable step sizes (Mohammadi et al., 2014). Another combines Distributed Economic Dispatch, Distributed State Estimation, an AR(2) load predictor, and a correction-plus-penalty mechanism based on pseudoinverse projections so that line overflows can be predicted and counteracted without any centralized facility (Li et al., 2019). In dynamic environments with unknown time-varying costs, a modified distributed primal–dual method framed as online convex optimization attains sub-linear static regret and sub-linear accumulated equality-constraint violation on the IEEE 14-bus system (Chatterjee et al., 2022).

More speculative computational work reformulates DC-OPF as a linearly constrained quadratic program inside a quantum interior-point method. The proposed noise-tolerant variant preserves primal and dual feasibility while allowing inexactness only in complementarity, and a classically augmented hybrid switches to a classical IPM when convergence slows (Amani et al., 2023). A plausible implication is that DC-OPF’s sparse KKT structure makes it a natural testbed for new linear-system technologies, even when hardware maturity is still limited.

5. Learning, differentiable optimization, and surrogate solvers

Learning-based DC-OPF methods exploit the fact that, for a fixed network, the map from load to optimal dispatch is structured rather than arbitrary. One statistical-learning approach observes that only a few active-set bases are relevant in most operating regimes, even though the total number of possible bases is exponential. It therefore learns the most probable bases from samples and evaluates an ensemble of basis-induced affine policies online. For most systems, ensemble policies built from as few as ten bases were reported to obtain optimal solutions with high probability, and with $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$2 the ensemble returned optimal solutions for $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$3 of scenarios in most tested systems (Ng et al., 2018).

Direct neural surrogates learn the DC-OPF map end-to-end. DeepOPF predicts normalized generator scaling factors $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$4, reconstructs angles through the reduced susceptance inverse,

$$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$5

and uses a convex post-processing projection when needed. On IEEE 30-, 57-, 118-, and 300-bus cases, it reported $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$6 feasibility in the test sets and two orders of magnitude speedup over conventional solvers (Pan et al., 2019). DeepOPF+ adds preventive calibration of line and slack-generator limits during training; its theoretical calibration uses PTDF row norms to bound how generator prediction error transfers to line-flow and slack-generation error, and the calibrated model reported $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$7 feasible solutions with minor optimality loss and two orders of magnitude speedup (Zhao et al., 2020).

A different strand embeds DC-OPF itself as a differentiable layer. One framework learns adjusted nodal shunt conductances $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$8 and branch susceptances $$f_\ell = \frac{1}{x_\ell}(\theta_i-\theta_j),\qquad -F_\ell \le f_\ell \le F_\ell,$$9 so that the modified DC-OPF more closely mimics AC-OPF, with gradients obtained by implicit differentiation of the KKT system through a differentiable solver (Rosemberg et al., 22 Mar 2025). Another performs offline parameter optimization of the DC approximation itself, tuning line coefficients and bias terms $p_g^{\min} \le p_g \le p_g^{\max},$0 by Truncated Newton Conjugate-Gradient so that DC-OPF generator setpoints better match AC-OPF over a training set; the reported improvements reached up to $p_g^{\min} \le p_g \le p_g^{\max},$1 in squared two-norm loss and $p_g^{\min} \le p_g \le p_g^{\max},$2 in $p_g^{\min} \le p_g \le p_g^{\max},$3-norm loss relative to traditional DC-OPF parameterizations (Taheri et al., 2024).

Recent hybrid architectures use DC-OPF as a fast feasible baseline and then refine it. Residual-learning models for AC-OPF take DC-OPF solutions as initialization and learn only the nonlinear corrections, reporting around $p_g^{\min} \le p_g \le p_g^{\max},$4 lower MSE, up to $p_g^{\min} \le p_g \le p_g^{\max},$5 lower feasibility error, and up to $p_g^{\min} \le p_g \le p_g^{\max},$6 runtime speedup relative to AC-IPOPT on OPFData benchmarks (Za'ter et al., 17 Oct 2025). For DC-OPF itself, a two-stage framework combining a physics-informed GNN with Continuous Flow Matching refines feasible initial dispatches toward near-optimality; on the IEEE 30-bus system it reported cost gaps below $p_g^{\min} \le p_g \le p_g^{\max},$7 for nominal loads, below $p_g^{\min} \le p_g \le p_g^{\max},$8 for extreme conditions, and $p_g^{\min} \le p_g \le p_g^{\max},$9 feasibility (Khanal, 11 Dec 2025).

6. Limitations, AC-feasibility gap, and data release

A recurring misconception is that DC-OPF solutions are merely approximate AC-OPF solutions and may sometimes be exactly AC-feasible. Under the assumptions used in one formal analysis—constant $\theta_{\mathrm{ref}}=0$0 loads, positive line resistances and reactances, symmetric admittance matrix, and at least one nonzero line flow so that AC losses are present—the projected feasible-generation sets satisfy

$\theta_{\mathrm{ref}}=0$1

Equivalently, no standard DC-OPF dispatch is AC-feasible, and even adding loss adjustments does not repair the mismatch because voltage magnitudes, reactive power, and nonlinear $\theta_{\mathrm{ref}}=0$2 and $\theta_{\mathrm{ref}}=0$3 terms remain absent from the DC model (Baker, 2019).

This does not make DC-OPF obsolete. Rather, it clarifies its role as a computational proxy, screening model, market-clearing surrogate, or baseline inside more accurate pipelines. The prominence of residual-correction models, parameter-tuned DC approximations, and differentiable DC layers suggests that the field increasingly treats DC-OPF as a structured first approximation whose output can be corrected, certified, or embedded in larger optimization stacks (Rosemberg et al., 22 Mar 2025, Za'ter et al., 17 Oct 2025).

A separate limitation concerns data release rather than dispatch quality. When aggregated DC-OPF outputs are published, privacy guarantees depend on network topology and on a monotonicity property of the OPF operator. For regional aggregation queries, monotone systems require Laplace noise with scale $\theta_{\mathrm{ref}}=0$4, whereas $\theta_{\mathrm{ref}}=0$5-monotone systems require scale $\theta_{\mathrm{ref}}=0$6. The same work shows that tree networks are monotone, while cyclic networks can be much farther from monotonicity, implying larger privacy noise for the same guarantee (Zhou et al., 2019).

Taken together, these results place DC-OPF in a precise technical position. It is a convex, highly structured, and operationally central optimization model whose algebraic simplicity enables strong duality, explicit active-set analysis, scalable GPU and distributed algorithms, differentiable layers, and feasible learning-based surrogates. At the same time, its lossless and active-power-only approximation imposes a clear AC-feasibility boundary that current research addresses through stochastic extensions, security constraints, parameter tuning, residual correction, and topology-aware privacy mechanisms (Rafiei et al., 2024, Baker, 2019).