Dispatch-Aware Deep Neural Network (DA-DNN)

• DA-DNN is an unsupervised framework that couples discrete topology decisions with continuous dispatch via a differentiable DC-OPF layer.
• It replaces the NP-hard combinatorial search in DC-OTS with a learned topology proposal, using generation cost as the training loss.
• Empirical evaluations on IEEE systems show that DA-DNN achieves near-optimal generation costs with runtime comparable to a single DC-OPF solve.

Dispatch-Aware Deep Neural Network (DA-DNN) denotes an unsupervised, optimization-in-the-loop framework for accelerating DC Optimal Transmission Switching (DC-OTS) by coupling discrete topology decisions with continuous dispatch through an embedded differentiable DC-OPF layer. In the formulation introduced for power systems, a neural line-switching module predicts relaxed line-status variables, these predictions gate line susceptances and thermal limits inside a convex DC-OPF, and the resulting generation cost is used directly as the training loss, so that physical constraints are enforced throughout training and inference without relying on presolved labels (Kim et al., 23 Jul 2025). In this setting, DA-DNN targets the computational bottleneck of mixed-integer DC-OTS, which is NP-hard, while retaining feasibility by construction and making inference cost the same as solving a single DC-OPF (Kim et al., 23 Jul 2025).

1. Problem setting and mathematical formulation

DA-DNN is defined in the context of optimal transmission switching, where a power network consists of buses $i \in \mathcal{B}$, lines $(i,j)\in\mathcal{E}$, and generators $g\in\mathcal{G}$. The principal variables are generator outputs $P_g$, bus voltage phase angles $\theta_i$, DC line flows $f_{ij}$, and binary line-status variables $z_{ij}\in\{0,1\}$, where $1$ denotes an in-service line and $0$ denotes an opened line. The baseline DC-OPF is a convex quadratic program over $\{P_g,\theta_i\}$ with quadratic generation cost, nodal power balance, generator limits, DC flow equations, thermal limits, angle bounds, and a slack-bus reference (Kim et al., 23 Jul 2025).

DC-OTS extends this baseline by optimizing topology through the binary variables $(i,j)\in\mathcal{E}$0. Its objective remains the generation cost,

$(i,j)\in\mathcal{E}$1

but power-balance and flow constraints become topology-dependent:

$(i,j)\in\mathcal{E}$2

$(i,j)\in\mathcal{E}$3

In matrix form, with branch-bus incidence matrix $(i,j)\in\mathcal{E}$4 and line susceptances vector $(i,j)\in\mathcal{E}$5,

$(i,j)\in\mathcal{E}$6

together with topology-scaled flow limits. Because this introduces mixed-integer structure into OPF, the problem becomes a mixed-integer quadratic program and is NP-hard (Kim et al., 23 Jul 2025).

The central motivation for DA-DNN is therefore not to replace OPF itself, but to amortize the combinatorial search over network topologies. A plausible implication is that the framework is best understood as a learned topology proposal mechanism whose outputs are always evaluated through a physical dispatch model rather than by direct surrogate prediction of dispatch variables.

2. Architecture and dispatch-aware mechanism

DA-DNN has two principal components. The first is a line switching layer $(i,j)\in\mathcal{E}$7, implemented as a multilayer perceptron, that ingests system inputs and outputs relaxed line-status variables $(i,j)\in\mathcal{E}$8. In the reported experiments, the input is the load vector $(i,j)\in\mathcal{E}$9, the MLP has 3 layers with ELU activations except at the final layer, and the hidden dimension is 64 for the IEEE 73-bus system and 128 for the IEEE 300-bus system (Kim et al., 23 Jul 2025). The output is passed through a sigmoid gate with scaling factor $g\in\mathcal{G}$0,

$g\in\mathcal{G}$1

which is used to reduce binarization error.

The second component is a differentiable DC-OPF layer. Given $g\in\mathcal{G}$2, the model constructs a gated network by defining

$g\in\mathcal{G}$3

and then solves a DC-OPF under these modified susceptances and limits. Dispatch awareness refers to the fact that neural topology predictions are not trained against target line labels; they are evaluated through the embedded OPF solver, and the generation cost returned by that solver becomes the learning signal (Kim et al., 23 Jul 2025).

This coupling is the distinguishing feature of the method. Instead of treating switching and dispatch as separable tasks, the network is trained only through the downstream economic consequences of its topology decisions. The formulation thereby couples discrete switching with continuous redispatch inside a single differentiable pipeline. Related work on multi-period DC-OPF adopts the same principle in a different setting: MPA-DNN projects unconstrained multi-period schedules onto a feasible dispatch set and trains directly on the projected operational cost, extending dispatch-aware learning from topology control to inter-temporal dispatch trajectories (Kim et al., 10 Oct 2025).

3. Differentiable DC-OPF layer and optimization-in-the-loop training

For a fixed relaxed topology $g\in\mathcal{G}$4, the embedded OPF layer solves a convex QP or LP over $g\in\mathcal{G}$5 with objective

$g\in\mathcal{G}$6

subject to equality and inequality constraints that are affine in both $g\in\mathcal{G}$7 and $g\in\mathcal{G}$8. The line parameters are masked as

$g\in\mathcal{G}$9

which preserves convexity of the DC-OPF subproblem during training (Kim et al., 23 Jul 2025).

The layer is differentiated through the KKT system. With dual variables $P_g$0 and $P_g$1 for equality and inequality constraints, the Lagrangian is

$P_g$2

At optimality, stationarity, primal feasibility, dual feasibility, and complementarity hold. Writing the stacked KKT equations as an implicit system $P_g$3, implicit differentiation gives the sensitivity of the optimal dispatch with respect to the relaxed topology:

$P_g$4

For network parameters $P_g$5 in $P_g$6, the training gradient follows by chain rule:

$P_g$7

In the original implementation, the OPF layer is solved inside the forward pass with CvxpyLayer using ECOS during training and Gurobi in testing (Kim et al., 23 Jul 2025). A later formulation reports CVXPYLayer with Clarabel during training and Gurobi during inference and metrics, while preserving the same OptNet-style implicit differentiation logic (Kim et al., 19 Dec 2025).

Continuous relaxation is used only for differentiability. During inference, $P_g$8 is thresholded, for example at $P_g$9, to obtain binary $\theta_i$0, and DC-OPF is re-solved to produce the final dispatch. The gating equations used in both training and testing remain

$\theta_i$1

so feasibility is attached to the OPF solve rather than to the raw neural outputs (Kim et al., 23 Jul 2025).

4. Loss design, feasibility-preserving initialization, and inference guarantees

The training loss is the generation cost evaluated at the OPF solution associated with the predicted topology:

$\theta_i$2

No presolved OTS labels are required, and in the reported experiments no extra penalties are used; the data are instead filtered for DC-OPF feasibility (Kim et al., 23 Jul 2025). The method can therefore be characterized as unsupervised in the sense that learning depends only on load scenarios and system parameters, not on optimal switching labels.

A critical practical feature is the feasibility-guaranteed initialization. The last layer of $\theta_i$3 is initialized as

$\theta_i$4

so that the first forward pass corresponds to the standard all-lines-in-service DC-OPF (Kim et al., 23 Jul 2025). Because the curated dataset is feasible under the base topology, the first OPF call is feasible. By contrast, random initialization scatters $\theta_i$5 in $\theta_i$6, frequently makes the embedded OPF infeasible, and blocks learning; the manual initialization is reported as essential for stable training on large grids (Kim et al., 23 Jul 2025).

Inference proceeds in three stages. First, given $\theta_i$7, the network computes $\theta_i$8. Second, it forms gated susceptances and thermal limits and solves DC-OPF to obtain $\theta_i$9 and the dispatch cost. Third, optionally, $f_{ij}$0 is thresholded to a binary topology and DC-OPF is re-solved. The final topology-dispatch pair is feasible by construction because it is the optimizer of a constrained DC-OPF under the realized topology (Kim et al., 23 Jul 2025). The runtime guarantee follows directly from this pipeline: once trained, DA-DNN requires one forward pass through $f_{ij}$1 and one DC-OPF solve, so the total wall-clock time equals that of a single DC-OPF (Kim et al., 23 Jul 2025).

A common misconception is to treat DA-DNN as a direct predictor of the optimal switching vector. The formulation does not support that interpretation. Its defining property is that switching decisions are meaningful only after they are filtered through the dispatch layer, and feasibility is guaranteed by the OPF solve rather than by the network alone.

5. Empirical performance, scalability, and robustness

The primary evaluation reported for DA-DNN uses PGLib-OPF IEEE 73-bus and 300-bus systems with 3,000 load snapshots generated under uniform 100–110% loading, discarding infeasible DC-OPF cases. Training uses AdamW with learning rate $f_{ij}$2, weight decay $f_{ij}$3, 50 epochs, and mini-batch size 25 (Kim et al., 23 Jul 2025).

On the IEEE 73-bus system with tight angle limits $f_{ij}$4, the reported average costs are DC-OPF $f_{ij}$5183.24\text{k}$f_{ij}$6\approx \$f_{ij}$7, and MILP DC-OTS $f_{ij}$8183.04\text{k}$f_{ij}$9\approx 5.23\text{ s}$z_{ij}\in\{0,1\}$0\overline{\theta}=0.4$z_{ij}\in\{0,1\}$1\approx \$z_{ij}\in\{0,1\}$2 with DC-OPF-level runtime (Kim et al., 23 Jul 2025). On the IEEE 300-bus system, DA-DNN achieves $z_{ij}\in\{0,1\}$3520.77\text{k}$z_{ij}\in\{0,1\}$4\approx \$z_{ij}\in\{0,1\}$5 at $z_{ij}\in\{0,1\}$6, a reduction of $z_{ij}\in\{0,1\}$7, while MILP DC-OTS is not solved within an hour; at $z_{ij}\in\{0,1\}$8, DA-DNN achieves $z_{ij}\in\{0,1\}$9514.34\text{k}$1$0\approx \$1$1, a reduction of $1$2, and MILP again times out (Kim et al., 23 Jul 2025).

A later study broadens the empirical picture to IEEE 73, 118, and 300-bus systems. It reports that on IEEE 118 in a binding regime, DA-DNN attains $1$3}$1$493.01\text{ k\$1$5 and DC-OPF $1$6}$1$719.14\pm199.44\text{ s}$1$8516.37\text{ k\$1$9 versus DC-OPF $0$0}$0$1\approx 1.55\%$ (Kim et al., 19 Dec 2025).

System / regime	Reported cost outcome	Computational outcome
IEEE 73, $0$2	DA-DNN $0$3183.10\text{k}$0$4\approx \$0$5	OTS $0$6 s per snapshot
IEEE 300, $0$7	DA-DNN $0$8520.77\text{k}$0$9\approx \$\{P_g,\theta_i\}$0	MILP not solved within an hour
IEEE 118, binding regime	DA-DNN $\{P_g,\theta_i\}$1}$\{P_g,\theta_i\}$293.01\text{ k\$\{P_g,\theta_i\}$3	OTS $\{P_g,\theta_i\}$4 s

The extended study also attributes a robustness property to the embedded OPF layer. When line flow limits are scaled to 90–130% at inference, training having been performed at 100%, DA-DNN is reported to track optimal OTS/DC-OPF behavior closely on IEEE 118, whereas a baseline without the OPF layer is insensitive and exhibits 100% violations (Kim et al., 19 Dec 2025). This suggests that dispatch-aware training yields not only speed and feasibility but also a degree of constraint-shift generalization, because the physical layer recomputes dispatch under the altered limits.

6. Relation to adjacent methods, limitations, and extensions

Within learning-based OPF and OTS, DA-DNN occupies a specific methodological niche. It differs from supervised learning approaches that require presolved OTS labels, because its loss is derived solely from the OPF solution cost. It also differs from penalty-based surrogates that predict dispatch or topology directly and penalize constraint violations, because DA-DNN embeds the constraint set inside the forward pass. A related line of work, MPA-DNN, applies the same dispatch-aware principle to multi-period DC-OPF with ramping constraints and storage dynamics by projecting unconstrained schedules onto a feasible multi-period set through a differentiable convex QP (Kim et al., 10 Oct 2025).

Several limitations are explicit. The framework currently embeds DC-OPF, so it inherits the DC approximation and does not model voltage magnitudes, reactive power, or losses. Security constraints such as $\{P_g,\theta_i\}$5 contingencies, remedial action constraints, and stability criteria are not included. Multi-period operations, interactions with unit commitment, and switching under uncertainty are deferred to future work. Connectivity constraints are also not structurally enforced; opening lines can create islands, and the text notes that connectivity or islanding constraints could be incorporated as additional constraints or penalties in the differentiable layer or via post-projection (Kim et al., 23 Jul 2025).

The discrete rounding step is another identified source of residual suboptimality. Thresholding $\{P_g,\theta_i\}$6 to binary $\{P_g,\theta_i\}$7 may introduce a small gap relative to the relaxed solution, even though the final dispatch remains feasible after re-solving OPF. Later discussion proposes straight-through estimators, bilevel rounding, safety filters, and hybridization with MILP solvers as possible extensions, including the use of DA-DNN to warm-start OTS or prune candidate lines (Kim et al., 19 Dec 2025).

The broader significance of DA-DNN is therefore not that it solves mixed-integer transmission switching exactly, but that it operationalizes a dispatch-aware learning paradigm: neural decisions are judged only through feasible downstream optimization. In power-system terms, this replaces explicit combinatorial search during inference with a learned topology generator whose outputs are economically and physically screened by OPF. This suggests a general template for optimization-in-the-loop control architectures in which feasibility is delegated to a differentiable solver and learning is driven by operational cost rather than imitation.