betaFeedback

WorksEngineering

Machine learning-guided real-time AC optimal transmission switching

Abstract

The operation of power grids has substantial economic consequences, affecting generation costs and, ultimately, electricity prices. The Optimal Transmission Switching (OTS) problem seeks to reduce these costs by selectively switching off transmission lines. However, OTS gives rise to a challenging mixed-integer nonlinear optimization problem that is difficult to solve at operationally relevant timescales, particularly as rapidly changing loads create a need for real-time switching decisions. To address this challenge, we develop a machine learning-guided method that identifies high quality line switching decisions in real time. Specifically, we construct a grid-specific switching-value model that rapidly estimates the generation-cost reduction associated with opening transmission lines by fine-tuning a pretrained model for AC optimal power flow (AC-OPF). The switching-value model provides signals to guide promising switching decisions, and topologies proposed by the model are evaluated using exact AC-OPF solvers to verify feasibility and obtain the actual generation cost. We develop multiple strategies that use the ML surrogate to efficiently explore the combinatorial space of transmission topologies, including Fixed-Rank Search, Adaptive-Rank Search, and Swap-Enhanced Search. Across multiple benchmark networks, our proposed algorithms identify line switching decisions that significantly reduce generation cost within a 50s computational budget. In particular, under a realistic data center buildout scenario on the 2,000-bus synthetic Texas grid, our approach reduces generation costs by 31.97% and gross nodal load payments by 94.34%. More broadly, our results demonstrate that a fine-tuned surrogate can effectively guide search even without guaranteeing feasibility, enabling classical optimization methods to tackle large nonlinear combinatorial problems on operationally relevant timescales.

Introduction

The cost and reliability of the electric grid have taken on a renewed importance around the world. In developed economies, electricity prices are now a central cost-of-living issue[1, 2, 3]. In other regions, the ability to provide access to electric power is fundamental for health and economic development[4, 5]. At the same time, the cost of electricity has increased at a rate that is much faster than inflation[6, 7].

There are many intertwined causes of the increase in electricity costs, but one of the main drivers is the dramatic increase in load. For example, in the US, approximately 10 GW of data center capacity entered the system in 2025[8]. Interestingly, generation in the US has kept up with load, with a total of 13.7 GW net additional capacity deployed in 2025, most of which is distributed solar and storage[9]. The part of the grid that has not changed is the transmission infrastructure. In 2025, the continental US constructed less than 900 miles of high voltage transmission lines [10]. To put this number in perspective, it is less than 0.3% of the US’s transmission infrastructure [11]. Therefore, there is a severe misalignment between rapid changes in load and generation and stagnation in transmission.

Addressing the root cause of this misalignment by expanding transmission infrastructure would be ideal, but it would take years, if not decades [12, 13]. In this paper, we propose an algorithmic framework that leverages the flexibility of the transmission system to achieve significant savings using existing infrastructure. More precisely, by judiciously switching off a small number of transmission lines, we can alleviate congestion and reduce the cost and price of electricity throughout the system. This joint optimization of network topology and power generation is known as the Optimal Transmission Switching (OTS) problem[14, 15], but its high computational cost has limited its adoption in routine operations. Building on a recently developed foundation model for power system operations and insights from power engineering [16, 17], we present an ML-guided framework that addresses these computational bottlenecks.

The need for faster and more reliable algorithms has always been a key impetus for power system research and engineering. System operators routinely solve problems such as unit commitment, economic dispatch and optimal power flow that determine how electricity should be generated and transmitted through large-scale grids [18]. OTS makes the optimization problem substantially harder by introducing the network topology itself as a decision: line-switching configurations create a combinatorial search, while each configuration must satisfy the nonconvex alternating current (AC) power flow equations. In recent years, advances in artificial intelligence (AI) and machine learning (ML) have created new opportunities to address this computational bottleneck, as learned models can exploit shared patterns across related optimization problem instances to guide algorithmic decisions [19, 20, 21].

However, state-of-the-art ML models are not, by themselves, reliable optimizers for complex decision-making problems in power systems. Operating points in power grids must satisfy physical laws and engineering constraints, and in general there are no hard guarantees on the reliability of ML predictions in realistic operating scenarios. Therefore, the main body of work in power systems uses ML models in conjunction with classical methods[22, 23, 24, 25, 26], aiming to achieve the speedup of ML while keeping the intuition and guarantees of existing methods, instead of fully replacing optimization with ML.

In this paper, we apply this general approach to the AC-OTS problem. OTS in AC admits a Mixed-Integer Nonlinear Programming (MINLP) formulation and can, in principle, be solved with general-purpose MINLP solvers. Direct MINLP and relaxation-based approaches have been successful on small- and medium-scale benchmark systems[27, 28]. However, AC-OTS combines two difficult sources of complexity: the continuous AC power-flow problem is nonlinear and nonconvex, while the line-status decisions introduce a combinatorial search over network topologies. This combination makes scaling to realistic transmission networks challenging. The need for fast solution methods is further amplified by increasingly dynamic operating conditions. Large computing loads, in particular, can exhibit substantial demand changes over short timescales[29, 30]. Recent ML-based approaches have also been shown to accelerate switching decisions in small- and medium-scale networks[23, 24, 25]. However, scaling these methods to larger systems might still incur substantial offline computational costs. In particular, training AC-OTS-specific models may itself require repeated expensive AC-OTS solves, for example through large numbers of episodes in reinforcement learning[24] or in the generation of AC-OTS solutions for supervised learning[23, 25]. Thus, although learning can reduce online solution time, the computational burden of building the ML model can become a bottleneck as network size increases.

These limitations motivate the two design principles of our ML-guided AC-OTS framework. First, the framework should be able to adapt and scale to large networks without requiring expensive AC-OTS-specific training from scratch. Second, the learned model should provide useful signals for navigating the combinatorial search, while the feasibility and the operating cost should be verified with exact solvers. This separation makes the framework robust to imperfect learned predictions without relying on them for correctness.

Guided by these principles, our framework consists of three components. First, we construct a grid specific switching value model, Line-switching Ranking Proxy, by adapting GridSFM[16, 17], a pretrained AC-OPF surrogate, to topology perturbations of each target network. Because fixing the switched lines reduces AC-OTS to AC-OPF on the resulting topology, GridSFM provides a natural foundation for estimating which switching actions are most effective in terms of generation cost. Second, we use insights from existing OTS heuristics to restrict the search to a set of candidate lines that are most promising for reducing operating costs. Third, we embed Line-switching Ranking Proxy within a learning-guided search that ranks switching moves over this candidate set and selectively evaluates the most promising switching actions with an AC-OPF solver. In this way, learning guides where to search, while AC-OPF solvers verify the feasibility and true operating costs of the resulting switching actions. We develop three variants of this approach, Fixed-Rank Search, Adaptive-Rank Search, and Swap-Enhanced Search, which differ in how they incorporate the surrogate predictions to explore the topology space, offering different tradeoffs between search flexibility and computational cost.

Using existing benchmarks and realistic data center build-out scenarios in Texas, we demonstrate that our proposed methods can efficiently reduce the cost by optimizing line switching and operation decisions of existing transmission infrastructure within a tight computational budget of 50 seconds. In particular, for the 2,000-bus Texas network with approximately 9 GW of added data center load, our best algorithm reduces generation cost by 31.97% and total load payments by 94.34% by switching only five transmission lines. These results demonstrate that even an imperfect ML surrogate can effectively guide search in a challenging nonlinear combinatorial space, bringing AC transmission switching closer to real-time operational use. All code, case-construction scripts, data and baselines will be released upon acceptance.

More broadly, AC-OTS is representative of a wider class of optimization problems in power systems in which discrete decisions are coupled with expensive nonlinear AC network optimization. The same structure appears in problems involving contingency analysis[31] and AC expansion planning[32]. Our results suggest a general paradigm for tackling such problems: use ML to capture useful structure in the underlying optimization landscape and to provide signals about where to search, rather than to replace optimization itself. Exact optimization methods can then retain responsibility for feasibility and solution quality. This framework could provide a path towards solving increasingly complex grid-control and planning problems at operationally relevant scales.

Background

We begin by motivating transmission switching as a mechanism for reducing generation costs. We then formulate the AC optimal transmission switching (AC-OTS) problem. We also review the standard AC optimal power flow (AC-OPF) problem, a key building block of AC-OTS providing the underlying power-flow optimization model when the transmission switching decisions are fixed.

Line Switching

A major part of electricity prices is transmission congestion. When a transmission line reaches its thermal limit, generation must be redispatched to respect the binding network constraint, potentially requiring more expensive generators and creating spatial differences in prices. Switching off a subset of transmission lines can redistribute flows around such constraints and restore access to lower-cost generation[14, 33]. This effect is counterintuitive because removing a transmission asset can improve, rather than degrade, the economic operation of a grid. Figure 1 illustrates this phenomenon: opening a transmission line changes the distribution of network flows and can yield a lower-cost dispatch than the original all-lines-in-service topology.

Three transmission network topologies illustrating line switching

Figure 1. This example shows that switching off different lines is optimal for different load profiles. In all three systems, the total load is 100 MW. However, different topology would optimize generation cost for different bus-level load profiles. In particular, for the middle and right networks, the base topology is not optimal. Rather, disconnecting lines changes the congestion patterns, and lead to lower costs.

AC Optimal Power Flow

For a fixed network topology and load profile, the AC-OPF problem determines how much active and reactive power each generator should produce, together with the voltage at each bus, so as to minimize generation cost while satisfying the nonlinear AC power-flow equations and operational constraints. This problem is nonconvex and has been a central object of study in power system operations[34, 35, 36]. Much progress has been made on solving it efficiently, but the problem remains difficult for large systems with thousands of buses, and solving one problem fully could take tens of seconds to minutes[37, 38, 39]. Since AC-OPF is a basic building block of many other decision making tasks in power system operations, algorithms that provide meaningful speedups would advance a host of problems.

We review the standard polar-coordinate formulation of the AC-OPF problem[18, 34]. We represent the transmission network by a set of buses N\mathcal{N} and a set of transmission lines E\mathcal{E}, with G⊆N\mathcal{G} \subseteq\mathcal{N} denoting buses with controllable generation. Each bus i∈Ni \in\mathcal{N} has a fixed active and reactive demand, denoted by pidp_i^{\mathrm{d}} and qidq_i^{\mathrm{d}}, respectively. The complex voltage at bus ii is represented in polar coordinates by its magnitude viv_i and phase angle θi\theta_i. At a generator bus i∈Gi \in\mathcal{G}, pigp_i^{\mathrm{g}} and qigq_i^{\mathrm{g}} denote the active- and reactive-power outputs of the generator; for buses without controllable generation, these quantities are zero. A transmission line between bus ii and bus kk is modeled as a complex admittance Gik+jBikG_{ik} + jB_{ik}, where GikG_{ik} is called the conductance, BikB_{ik} the susceptance, and jj the imaginary unit. We collect bus voltages, bus angles, generator active power and generator reactive powers into vectors, denoted by v\mathbf{v}, θ\boldsymbol{\theta}, p\mathbf{p}, and q\mathbf{q}, respectively. The AC-OPF minimizes generation cost subject to the power flow equation and engineering constraints:

min⁡v,θ,p,q∑i∈Gc2,i(pig)2+c1,ipig+c0,i(1)\begin{aligned} \min_{\mathbf{v},\boldsymbol{\theta},\mathbf{p},\mathbf{q}}\quad&\sum_{i\in\mathcal{G}} c_{2,i}(p_i^{\mathrm{g}})^2+c_{1,i}p_i^{\mathrm{g}}+c_{0,i} \tag*{(1)} \end{aligned}
pig−pid=vi∑k∼ivk(Gikcos⁡θik+Biksin⁡θik),i∈N,(2)p_i^{\mathrm{g}}-p_i^{\mathrm{d}}=v_i\sum_{k\sim i}v_k(G_{ik}\cos\theta_{ik}+B_{ik}\sin\theta_{ik}),\qquad i\in\mathcal{N}, \tag*{(2)}
qig−qid=vi∑k∼ivk(Giksin⁡θik−Bikcos⁡θik),i∈N,(3)q_i^{\mathrm{g}}-q_i^{\mathrm{d}}=v_i\sum_{k\sim i}v_k(G_{ik}\sin\theta_{ik}-B_{ik}\cos\theta_{ik}),\qquad i\in\mathcal{N}, \tag*{(3)}
pig≤pig≤p‾ig,i∈G,(4)p_i^{\mathrm{g}}\leq p_i^{\mathrm{g}}\leq\overline{p}_i^{\mathrm{g}},\qquad i\in\mathcal{G}, \tag*{(4)}
qig≤qig≤q‾ig,i∈G,(5)q_i^{\mathrm{g}}\leq q_i^{\mathrm{g}}\leq\overline{q}_i^{\mathrm{g}},\qquad i\in\mathcal{G}, \tag*{(5)}
v‾i≤vi≤v‾i,i∈N,(6)\underline{v}_i\leq v_i\leq\overline{v}_i,\qquad i\in\mathcal{N}, \tag*{(6)}
[vivk(Gikcos⁡θik+Biksin⁡θik)]2+[vivk(Giksin⁡θik−Bikcos⁡θik)]2≤S‾ℓ 2,(i,k)∈E,(7)\left[v_iv_k(G_{ik}\cos\theta_{ik}+B_{ik}\sin\theta_{ik})\right]^2+\left[v_iv_k(G_{ik}\sin\theta_{ik}-B_{ik}\cos\theta_{ik})\right]^2\leq\overline{S}_{\ell}^{\,2},\qquad(i,k)\in\mathcal{E}, \tag*{(7)}

where we write θik=θi−θk\theta_{ik}=\theta_i-\theta_k and k∼ik\sim i when bus kk is connected to bus ii. In the objective we adopt the commonly used quadratic cost functions, but other cost functions can be used as well. The active- and reactive-power balance constraints (2) and (3) ensure that, at each bus, local generation minus demand equals the net power injection determined by the AC power-flow equations. Thus, demand that cannot be met by local generation must be supplied through the network, with total generation covering both demand and transmission losses. The remaining constraints impose generator active- and reactive-power limits (4) and (5), voltage magnitude limits (6), and transmission-line flow limits (7).

We can write (1) to (7) in a more compact form as

min⁡x∑i∈Gc2,i(pig)2+c1,ipig+c0,is.t.x∈FAC,(8)\min_{\mathbf{x}}\quad\sum_{i\in\mathcal{G}} c_{2,i}(p_i^{\mathrm{g}})^2+c_{1,i}p_i^{\mathrm{g}}+c_{0,i}\qquad\text{s.t.}\quad\mathbf{x}\in\mathcal{F}_{\mathrm{AC}}, \tag*{(8)}

where the objective minimizes the generation cost over the variables x=(v,θ,p,q)\mathbf{x}=(\mathbf{v},\boldsymbol{\theta},\mathbf{p},\mathbf{q}) and FAC\mathcal{F}_{\mathrm{AC}} denotes the feasible set.

AC Optimal Transmission Switching

The AC optimal transmission switching (AC-OTS) problem extends AC-OPF by jointly optimizing power-system operation and transmission topology. [14] It determines which transmission lines should be in service, in addition to deciding generator outputs and bus voltages, so as to minimize generation cost subject to the AC power-flow equations and operational constraints.1 For notational simplicity, we formulate the problem in terms of switching lines off from a base topology, although our method applies equally when some lines can be switched on from a nominally off position.

Formally, the decision variables are S\mathcal{S} and x\mathbf{x}, where S⊆E\mathcal{S}\subseteq\mathcal{E} is the set of transmission lines to be opened and x\mathbf{x} collects all the AC-OPF decision variables introduced above. We denote by $\mathcal{F}_{\mathrm{AC}}(\mathcal{S}) the set of AC-OPF solutions x\mathbf{x} that are feasible for the network obtained by opening the lines in SS. The AC-OTS problem can then be written compactly as

min⁡S⊆E, x∑i∈Gc2,i(pig)2+c1,ipig+c0,is.t.x∈FAC(S),∣S∣≤J.(9)\begin{aligned} \min_{S\subseteq E,\ \mathbf{x}}\quad&\sum_{i\in\mathcal{G}} c_{2,i}(p_i^{g})^2+c_{1,i}p_i^{g}+c_{0,i}\\ \text{s.t.}\quad&\mathbf{x}\in\mathcal{F}_{\mathrm{AC}}(S),\\ &|S|\le J. \tag*{(9)} \end{aligned}

Here, decisions are both the continuous AC power flow variables x\mathbf{x} and the discrete topological variables SS. Hence, the generation costs are coupled with the selected transmission topology. This formulation includes the cardinality constraint ∣S∣≤J|S|\le J on the maximum number of lines that can be opened, which limits the extent to which the optimized topology can deviate from the base network. This constraint reflects the fact that AC-OPF only models the steady state behavior of the grid. Dynamic stability of the system is typically checked beforehand for the base topology and some of its close variants. Therefore, JJ should not be too large. In addition, we show that most of the benefit in line switching is captured by a small number of switches.

We remark that AC-OTS is significantly more challenging to solve than the already difficult AC-OPF problem. On networks with thousands of buses and transmission lines, state-of-the-art optimization solvers struggle to find even an improving solution, despite the existence of topologies with substantially lower generation cost [41, 42].

Prior Approaches to AC-OTS

Since direct mixed-integer formulations of AC-OTS are not computationally tractable for medium-to large-scale grids, substantial work has focused on methods for finding high-quality switching decisions within practical time limits. These approaches broadly fall into three categories: approximations using DC power flow, AC-informed approximations, and learning-based methods that directly target the nonlinear AC-OTS problem.

The first class of approaches relies on DC approximations, replacing the nonlinear AC power-flow equations with a tractable linear model for identifying promising switching actions. Early work uses DC-OPF solutions to rank candidate lines [15], while more recent methods leverage machine learning to prioritize candidate switches [43], predict switching topologies [22, 23], or learn topology decisions through differentiable DC-OPF layers [44]. Although computationally attractive, these approaches inherit the limitations of the DC approximation. More precisely, switching decisions that appear promising may actually be infeasible for the full AC model [45]. Therefore, a list of candidates can often contain no good or even feasible actions for the AC-OTS problem.

A second line of work seeks to retain the tractability of DC-OTS while incorporating information from the AC model. Feasibility can be improved by calibrating DC-model parameters using AC-OPF information [46], or augmenting the DC-OTS formulation with AC admittance sensitivities and repair costs that approximate the effects of topology changes on AC operation [47].

Finally, learning can also be applied directly to nonlinear topology optimization. A network- and objective-specific line-selection policy was trained with reinforcement learning [24], and constraint-aware diffusion was used to predict complete switching configurations [25]. These approaches, however, require substantial offline computation, often needing to generate tens of thousands of labels per network [24, 25]. This computational burden becomes increasingly significant as networks grow in size. The largest AC-OTS benchmark reported contains 500 buses [25], falling far short of the scale of large transmission networks of practical interest.

These limitations motivate methods that can search directly over AC-feasible topology changes without relying on expensive offline training.

Methods

We develop a learning-guided system for real-time AC-OTS that uses fast, learned estimates to explore the combinatorial space of switching actions while reserving full AC-OPF solves (using conventional solvers) for evaluating promising candidate topologies.

Our system is guided by two design principles. The first is the ability to adapt and scale to large systems. Rather than learning a network-specific switching policy from scratch, we provide a scalable recipe for constructing Line-switching Ranking Proxy for each target grid. We begin with GridSFM [16, 17], a pretrained model for estimating AC-OPF solutions, and adapt it through a task-specific contrastive fine-tuning procedure. The resulting model is trained specifically to distinguish switching actions according to their effect on generation cost, rather than to perform generic cost prediction. The second is that the ML model and conventional solvers play distinct roles. The former provides guidance and the latter provide verification. We use Line-switching Ranking Proxy to guide the combinatorial search and we do not require its estimates to guarantee feasibility or improvement in the true AC objective. Instead, Line-switching Ranking Proxy must provide a sufficiently informative signal to identify promising moves, which are then evaluated using AC-OPF solves to determine their feasibility and true operating cost. This separation makes the system robust to imperfect learned estimates while retaining solver-based verification of every accepted topology.

Based on these principles, our system comprises three components. First, we construct a grid-specific switching-value model, denoted by Line-switching Ranking Proxy (§ 3.1), that rapidly estimates the generation-cost reduction associated with opening transmission lines. Second, we use domain knowledge from existing OTS heuristics to identify a set of eligible lines that are most relevant to congestion relief (§ 3.2). Third, we embed Line-switching Ranking Proxy within learning-guided local search algorithms that rank moves over the resulting neighborhoods and selectively evaluate the most promising candidates using an AC-OPF solver (§ 3.3).

The three components jointly balance search breadth and computational cost in real-time AC-OTS. Although Line-switching Ranking Proxy is substantially faster than AC-OPF, scoring every possible switching action remains too expensive within a real-time budget. Domain-informed filtering therefore narrows the search to lines most likely to affect congestion, keeping ML inference tractable. At the same time, Line-switching Ranking Proxy allows us to retain a much broader set of candidate lines than would be practical with AC-OPF evaluations alone. Within this larger search space, learning-guided local search rapidly prioritizes promising moves and reserves expensive AC-OPF solves for selective verification. Together, these components enable broader yet computationally efficient exploration while ensuring that every accepted topology is verified for feasibility and operating cost.

Grid-Specific Switching-Value Model

GridSFM

GridSFM [16, 17] is a graph transformer model that estimates solutions to the AC-OPF problem. The model has approximately 15 million parameters and operates directly on the structure of the transmission network. It represents a power system as a heterogeneous graph comprising buses, generators, loads, shunts, transmission lines, transformers, and network cycles, with edges encoding the physical relationships among these components. Network cycles—closed loops of transmission lines and transformers—are represented explicitly, allowing the model to capture coupling among line flows induced by Kirchhoff’s voltage law. The model consists of blocks that combine linear attention among components of the same type with local message passing between physically connected components, thereby capturing interactions in the grid. Its output heads predict AC-OPF decision variables, including generator active-power and voltage setpoints; a subsequent power-flow completion step recovers the remaining electrical state, from which generation cost can be evaluated.

Importantly for our setting, GridSFM is pretrained jointly across a diverse collection of transmission networks and operating conditions rather than trained separately for a single grid. The model is therefore exposed during pretraining to variations in both operating conditions and network topology, allowing it to learn transferable relationships between network structure, power flow and generation cost. This is particularly relevant to AC-OTS, since once a set of lines S⊆ES \subseteq\mathcal{E} is opened, evaluating the resulting topology amounts to solving an AC-OPF problem on the modified network. GridSFM thus provides a natural pretrained starting point for rapidly assessing the effect of candidate topology changes.

GridSFM can be adapted to a target network using comparatively few solved AC-OPF instances, making it a natural foundation for constructing our grid-specific Line-switching Ranking Proxy. Our objective, however, is not generic AC-OPF prediction. Instead, we require Line-switching Ranking Proxy to discriminate among switching actions according to their effect on generation cost. We therefore introduce a task-specific contrastive fine-tuning procedure for this purpose, described next. We refer the reader to Bhan et al. [17] for complete details of the GridSFM architecture and training.

Fine-tuning pipeline

GridSFM is fine-tuned to capture how local topology changes affect the AC-OPF objective within a given grid. We create training data from both load perturbation and topology perturbation for each grid. For each target grid, we generate multiple load-perturbed operating scenarios. For each scenario, we construct sub-networks by removing one line, for all the switchable lines in the network, as illustrated in Figure 2, discarding removals that disconnect the network. We solve AC-OPF on each remaining topology to obtain labels for fine-tuning. We designed the data generation pipeline so that the number of labeled sub-networks grows linearly with the number of switchable lines in the network, for a fixed number of load scenarios. To keep offline data collection scalable, we avoid generating combinatorially many multi-line switching sub-networks, while still exposing Line-switching Ranking Proxy to single-line changes that serve as the basic moves in our subsequent search procedure.

Figure 2. Fine-tuning pipeline to build the Line-switching Ranking Proxy. Training data combines load perturbations with single-line topology perturbations. Connectivity-preserving sub-networks are solved with exact solvers to obtain AC-OPF solutions for fine-tuning. Pretrained GridSFM is fine-tuned on sub-networks from each grid to obtain a grid-specific Line-switching Ranking Proxy.

Internally, GridSFM represents each transmission line as an edge between its endpoint buses, and associate it with features such as admittance values and thermal ratings. Line-switching Ranking Proxy augments the feature vector of a line by adding a binary in-service indicator. This indicator is set to one for every line in the base topology, and switching off a line is represented by setting its indicator to zero. Thus, different candidate switching configurations can be represented by modifying the in-service indicators while retaining the same architecture and model parameters.

Next, we augment the physics- and cost-based losses used by GridSFM with a contrastive ranking objective that directly promotes promising switching actions. For each load scenario, we rank the feasible single-line perturbations according to their AC-OPF generation cost and designate the KK lowest-cost candidates as positive examples, with the remaining candidates serving as negative examples. The resulting contrastive objective encourages Line-switching Ranking Proxy to distinguish promising switching actions from less valuable ones. The value of KK controls how wide a band of promising topologies the model is encouraged to identify and is treated as a hyperparameter.

The contrastive objective emphasizes the relative ordering of candidate topologies rather than requiring highly accurate absolute cost predictions. This distinction is important for the design of our search procedure: candidate topologies within the same operating scenario often have similar total costs, while the search uses only their relative ranking to decide which candidates should receive expensive AC-OPF verification. We therefore train the model to distinguish promising topology changes even when its absolute cost estimates are imperfect.

Although the same framework could in principle use any ML model that estimated AC-OPF solutions, the ability to generalize across network topologies is particularly important for AC-OTS. Many existing AC-OPF prediction models are trained on a fixed network topology while varying only the operating conditions [48, 49], whereas AC-OTS requires evaluating topologies that differ from the intact network. GridSFM is well suited to this setting because its pretraining spans transmission networks of different sizes and topologies, and its graph-based representation can naturally encode topology perturbations through the added binary in-service line indicator. Fine-tuning on single-line removals then adapts the pretrained model to the local topology changes encountered on each target network during AC-OTS search.

Candidate-Line Filtering

Although Line-switching Ranking Proxy can score candidate switching actions much faster than AC-OPF, evaluating every possible line switching remains impractical for real-time decision making on large networks. For example, using a Nvidia H100 GPU, it would take GridSFM about 1 minute to evaluate the cost of switching every single line for a grid with thousands of buses. While significantly faster than conventional solvers, it would still take more than the allocated time budget in real-time operations. We therefore first restrict the search to a set of candidate lines using two complementary filtering strategies from the OTS literature. The underlying observation is that topology impacts the cost through congestion.2 Thus, both strategies target lines that can affect congested parts of the network, either through topological proximity or through their estimated impact on congested flows.

Following the above intuition and existing OTS heuristics [15, 50], we create a congestion-region filter and retain lines in the vicinity of congested regions of the network. More precisely, we perform a breadth-first search extending two bus hops from both endpoints of each congested line, and every line incident to a visited bus is included in the candidate set. This filter captures nearby switching actions that may reroute power and change the congestion pattern[51].

Topological proximity to congested lines alone can miss useful switching actions. Through Kirchhoff’s voltage law, power flows are coupled along network cycles, which may span topologically distant parts of the grid. Consequently, opening a line far from a congested line can still substantially alter its flow. We account for these nonlocal effects using line-outage distribution factors (LODFs). These factors form a matrix, where the (i,k)(i,k)th element captures the sensitivity of the changes of power flows on line ii given a change on line kk[52]. We create a LODF-based filter by adding the 50 lines outside of the congested region based on the largest LODF values.

The resulting candidate set is deliberately much larger than those considered by existing OTS heuristics [51, 53]. We can afford this broader search space because we leverage the speed of Line-switching Ranking Proxy for a fast estimation of the value of moves. This extended exploration increases the opportunity to identify lower-cost switching configurations.

Learning-Guided Local Search

We now describe our learning-guided local-search framework, which combines the candidate-line filtering of § 3.2 with the fast switching-value estimates provided by Line-switching Ranking Proxy. We call a line eligible if it is retained by the filtering step, and denote the set of eligible lines by E\mathcal{E}. At any point in the search, the current network is represented by the set SS of opened lines, and a neighborhood consists of candidate moves that modify SS using only lines retained by the filtering step. Depending on the search variant, a move may open an additional line or simultaneously open a line and close a previously open one.

At each iteration, the local search alternates between two steps. Exploration. Candidate moves are ranked according to their estimated generation-cost benefit using Line-switching Ranking Proxy. Depending on the search variant, move values are either recomputed using the current topology by invoking Line-switching Ranking Proxy or approximated using estimates obtained from the base topology, without further model inference. The search considers moves in order of estimated value, starting with the most promising. Verification. A move is never accepted based on Line-switching Ranking Proxy alone. Before accepting a candidate, we solve the full AC-OPF problem on the resulting topology to verify both its feasibility and actual generation cost. If the candidate is feasible and improves upon the incumbent, the move is accepted; otherwise, the search proceeds to the next candidate.

Once a move is accepted, the resulting topology becomes the new incumbent and the exploration–verification process repeats. The search continues until no further improving move is identified, the maximum number JJ of opened lines is reached, or the prescribed time budget is exhausted.

We instantiate this learning-guided search in three variants that trade off the computational cost and quality of each search step, which we describe next; their pseudocodes are provided in the Appendix.

Fixed-Rank Search (FRS)

For any set of opened lines SS, let C^θ(S)\widehat{C}_{\theta}(S) denote the generation cost predicted by LINE-SWITCHING RANKING PROXY for the topology obtained by opening the lines in SS from the base network. Fixed-Rank Search (FRS) is our least computationally expensive variant per search step: it invokes LINE-SWITCHING RANKING PROXY only once, before the search begins, and reuses the resulting ranking throughout the search. Its workflow is illustrated in Figure 3.

Overview of the learning-guided search framework for Fixed-Rank Search

Figure 3. Overview of the learning-guided search framework, illustrated for Fixed-Rank Search. The eligible lines E′\mathcal{E}' are first scored by the learned line-switching ranking model to define an exploration order π\pi. At each iteration, the next candidate line is proposed and the resulting topology is verified using a full AC-OPF solve. The switching action is accepted only if it reduces the verified operating cost. This separation between learning-guided exploration and solver-based verification underlies all three proposed search variants.

Specifically, FRS evaluates the removal of each eligible line b∈E′b \in\mathcal{E}' from the base topology and orders the lines by non-decreasing predicted cost,

π=(b1,…,b∣E′∣),C^θ({b1})≤⋯≤C^θ({b∣E′∣}).\pi= (b_1,\ldots,b_{|\mathcal{E}'|}), \qquad\widehat{C}_{\theta}(\{b_1\}) \leq\cdots\leq\widehat{C}_{\theta}(\{b_{|\mathcal{E}'|}\}).

This ranking is computed once from the base topology and reused throughout the search, requiring no further LINE-SWITCHING RANKING PROXY inference as the topology evolves.

Starting from S=∅S = \emptyset, FRS considers lines sequentially according to π\pi. For each line bb, the candidate move is to open bb, producing the candidate set of opened lines S∪{b}S \cup\{b\}. If the move preserves network connectivity, it undergoes the AC-OPF verification step described above. If the resulting topology is feasible and improves the incumbent cost, the move is accepted and S←S∪{b}S \leftarrow S \cup\{b\}; otherwise, bb is rejected and is not reconsidered in the future. This step is repeated until one of the stopping criteria described above is met.

Adaptive-Rank Search (ARS)

Adaptive-Rank Search (ARS) adapts its exploration step to the switching decisions of the previous iterations of the search. Unlike FRS, it does not reuse ranking information. Indeed, at every iteration, ARS invokes LINE-SWITCHING RANKING PROXY to reevaluate all eligible opening moves from the current set SS of opened lines. Specifically, each eligible line b∈E′∖Sb \in E' \setminus S defines the move S→S∪{b}S \to S \cup\{b\}, whose predicted cost is

C^θ(S∪{b}).\widehat{C}_{\theta}(S \cup\{b\}).

ARS ranks these moves by non-decreasing predicted cost and applies the AC-OPF verification step in this order until an improving feasible move is identified.

If the move associated with opening line bb is accepted, we update S←S∪{b}S \leftarrow S \cup\{b\} and begin a new iteration, reevaluating all eligible opening moves from the updated set SS. Note that, unlike FRS, an eligible line whose opening was rejected at an earlier iteration may become promising after other lines have been opened and is therefore reconsidered. This process is repeated until one of the stopping criteria described above is met.

Swap-Enhanced Search (SES)

Swap-Enhanced Search (SES) expands the local-search neighborhood by allowing previous switching decisions to be undone. In addition to an opening move S→S∪{b}S \to S \cup\{b\}, SES considers swap moves that re-close a line a∈Sa \in S while opening an eligible line b∈E′∖Sb \in E' \setminus S, producing the set of opened lines (S∖{a})∪{b}(S \setminus\{a\}) \cup\{b\}.

SES has two phases. In the first phase, it computes the fixed ranking π\pi of FRS (§ 3.3.1) and follows this search algorithm until five improving line openings have been accepted or the line-opening limit JJ is reached. In the second-phase, it constructs two neighborhoods at each iteration. The opening neighborhood consists of moves that open one of the ten highest-ranked eligible lines in the fixed FRS ranking π\pi that remain closed. The swap neighborhood is formed by pairing each of these lines with each of the five most recently opened lines, yielding up to 5×10=505 \times10 = 50. SES invokes Line-switching Ranking Proxy to evaluate every move in these neighborhoods.

The two moves with lowest predicted cost from each neighborhood are selected for AC-OPF verification. If at least one of these moves is feasible and improves upon the incumbent, the move with lowest verified cost is accepted and SS is updated accordingly. The neighborhoods are then rebuilt around the updated set of opened lines, and Line-switching Ranking Proxy reevaluates all moves in the new neighborhoods.

Evaluation

Networks

We evaluate the proposed learning-guided methods on three transmission networks spanning a wide range of system sizes: the api variants of the IEEE 118-bus and Polish 2,737-bus networks from the PGLib-OPF benchmark library[54], and Texas2k, the 2,000-bus synthetic Texas grid (ACTIVSg2000, 2025-series summer peak)[55, 56]. The api variants of the IEEE 118-bus and Polish systems are stressed operating cases with elevated loading and are commonly used to evaluate AC-OPF and AC-OTS algorithms under congested conditions. We use these two networks to evaluate the performance of our proposed methods on established benchmarks. Texas2k does not have an analogous api variant. To construct a comparably stressed setting, we therefore introduce data center loads at locations corresponding to publicly announced projects. For Texas2k, we consider the following scenarios, and additional details on the construction of the Texas2k scenarios are provided in the Appendix. In the networks studied, every line whose removal does not island the network is considered switchable.

Texas2k–data centers. We construct a concentrated-load-growth scenario using publicly announced data center projects for which both location and planned capacity are available [57, 58, 59, 60]. The projects are mapped to geographically corresponding locations in Texas2k, adding approximately 9.1 GW of load in total, corresponding to a 10.7% increase relative to the original 85.8 GW summer-peak demand.

Texas2k–Fort Worth halved. Because announced data center capacity is uncertain and may not fully materialize, we additionally consider a lower-load realization in the Fort Worth hub. We reduce the added Fort Worth data center load to 50% of its value in the Texas2k–data centers scenario while leaving the added load at all other data center hubs unchanged.

Baselines

We compare our methods with six baselines: Price-difference sequential heuristic [15] (PSeq), Random Congestion (RC), Congestion Simulated Annealing (CSA), AC-OTS with Gurobi, and DC-OTS with Gurobi, and Juniper [61].

Price-difference sequential heuristic (PSeq). This is the sequential LP heuristic of Fuller et al. [15], adapted to the AC setting. Every in-service, non-islanding line kk is scored by the ranking parameter αk=(πto−πfrom)×Pk\alpha_k = (\pi_{to} - \pi_{from}) \times P_k, the price difference across the line times its active flow. A negative value marks counter-economic flow, from a higher-price bus to a lower-price one. We walk down the lines in order from the most negative αk\alpha_k. The first removal that lowers the AC-OPF cost is accepted.

Random Congestion (RC). RC uses the same two-hop congestion and top cycle set as in the GridSFM-guided methods. Lines in these two sets are listed in random order, and this order remains fixed throughout the execution of the method. RC proceeds in this order and accepts a line to open if opening it lowers the AC-OPF cost.

Congestion Simulated Annealing (CSA). CSA also uses the same two-hop congestion and top cycle set as in the GridSFM-guided methods. At each step it proposes a uniformly random move: open a line, reclose an opened line, or swap one for the other. Openings, including a swap’s opening, come from the same candidate pool as random order, while reclosures are unrestricted. Each proposal is solved exactly and accepted if it lowers the AC-OPF cost.

Gurobi AC-OTS. The exact optimal transmission switching problem in the polar AC formulation, built with PowerModels [62].

Gurobi DC-OTS. The same switching problem in the linear DC power-flow approximation, for which the solution may not be AC-feasible.

Juniper. The same AC optimal transmission switching model, solved by Juniper [61], an open-source nonlinear branch-and-bound solver. Juniper solves continuous relaxations with Ipopt and uses a feasibility pump with HiGHS to find integer-feasible solutions.

Metrics

Generation cost savings (%). We compare the methods using the percentage reduction in generation cost relative to the intact topology. Let SS denote the set of opened transmission lines found by a method, and let C(S)C(S) be the corresponding generation cost obtained from a full AC-OPF solve on the resulting topology. The generation-cost savings are defined as

GenerationSavings⁡(S)=100⋅C(∅)−C(S)C(∅),(10)\operatorname{Generation Savings}(S) = 100 \cdot\frac{C(\varnothing) - C(S)}{C(\varnothing)}, \tag*{(10)}

where C(∅)C(\varnothing) is the AC-OPF generation cost of the intact topology with all lines in service.

Load payment savings (%). For Texas2k, we additionally report the percentage reduction in total load payment, as the realistic data center scenarios make the resulting prices and load payments more interpretable. Specifically, every solution to AC-OTS also yields nodal prices from its dual variables[3]. The nodal price, also known as the locational marginal price, in $/MWh, is the dual variable of the active power balance constraint at bus ii, denoted by πi\pi_i[63]. It represents the price of supplying an additional unit of electricity demand at that location, given the current operating conditions. Let pidp_i^d denote the load served at bus ii in MW, and pkgp_k^g the dispatch of generator kk located at bus kk. The hourly settlements are then computed as

Total load payment: Π=∑iπipid,Total generator revenue: R=∑k∈Gπkpkg.(11)\text{Total load payment: } \Pi= \sum_i \pi_i p_i^d,\qquad\text{Total generator revenue: } R = \sum_{k\in G} \pi_k p_k^g. \tag*{(11)}

We note that the market price πi\pi_i represents only a part of the utility bill, the other being fixed costs related to infrastructure and upgrades. Nevertheless, the increase in the prices and load payments would eventually be passed on to the customers. We compare the methods using the percentage reduction in total load payment relative to the intact topology. Let Π(S)\Pi(S) denote the total load payment under the topology with opened-line set SS. The load-payment savings are defined as

PaymentSavings⁡(S)=100⋅Π(∅)−Π(S)Π(∅),(12)\operatorname{Payment Savings}(S) = 100 \cdot\frac{\Pi(\varnothing) - \Pi(S)}{\Pi(\varnothing)}, \tag*{(12)}

where Π(∅)\Pi(\varnothing) is the total load payment under the intact topology.

Computational Setup

The maximum number of lines that can be opened (JJ) is set to 1% of the total number of lines for each network. All methods use a time limit of 50s. We use MadNLP[64], a GPU-accelerated interior point algorithm-based solver to solve the AC-OPF problems. All GridSFM-guided methods are run on NVIDIA H100 GPUs. The Juniper, DC-OTS and AC-OTS baselines are run on AMD EPYC 7542 CPUs. The software versions are Gurobi v12.0.3, Juniper 0.9.4, Ipopt 1.16.0, and HiGHS 1.25.4. For algorithms that are randomized (RC and CSA), we report the results with 10 random seeds.

Results

Results

Benchmarks

The results for IEEE 118-bus-api and Polish 2,737-bus-api are shown in Table 1. On IEEE 118-bus-api, the proposed learning-guided Swap-Enhanced Search (SES) finds the optimal solution in terms of generation cost savings. On Polish 2,737-bus-api, both learning-guided FRS and SES outperform all other baselines. The solution found in 50 seconds is close to 50% of the upper bound on savings found by Juniper using a 24-hour run.

MethodIEEE 118-bus-api (+62.1 % load)Polish 2,737-bus-api (+84.0 % load)
Gen. savings (%)# LinesGen. savings (%)# Lines
Ours (50s)
FRS2.5912.6615
ARS2.5911.899
SES4.7412.417
Baselines (50s)
PSeq4.7412.108
RC0.79±1.090.79 \pm1.091.0±0.01.0 \pm0.01.69±0.411.69 \pm0.4120.4±2.120.4 \pm2.1
CSA3.08±1.323.08 \pm1.321.0±0.01.0 \pm0.00.94±0.520.94 \pm0.5217.6±8.517.6 \pm8.5
Juniper0.0000.000
Gurobi AC-OTS0.0000.000
Gurobi DC-OTS0.001†1^{\dagger}0.000
Reference (24h)
Juniper4.74 (opt)15.1826

Table 1. Generation-cost savings and number of opened lines for IEEE 118-bus-api and Polish 2,737-bus-api at a 50 s search budget. Compared to the nominal cases in PGLib, the api case increases the load by 62.1 % and 84.0 % in the IEEE 118-bus and Polish 2,737-bus networks, respectively. Seeded baselines report mean ±\pm standard deviation over 10 seeds. †: the topology obtained by Gurobi DC-OTS is not AC-feasible.

The results for the two Texas2k scenarios are shown in Table 2. The learning-guided methods outperform all other baselines in both scenarios. SES performs the best in both generation-cost savings and payment savings in both scenarios.

MethodTexas2k–data centers (+10.7 % load)Texas2k–Fort Worth halved (+9.9 % load)
Gen. savings (%)Payment savings (%)# LinesGen. savings (%)Payment savings (%)# Lines
Ours (50s)
FRS31.9494.31318.1488.274
ARS31.9494.31318.1488.274
SES31.9794.34518.1688.345
Baselines (50s)
PSeq23.7378.12114.0679.391
RC5.30±10.235.30 \pm10.2324.14±32.1124.14 \pm32.111.2±1.51.2 \pm1.58.66±6.438.66 \pm6.4341.16±30.9941.16 \pm30.994.8±3.04.8 \pm3.0
CSA0.51±0.720.51 \pm0.727.07±9.737.07 \pm9.730.9±0.60.9 \pm0.63.53±5.413.53 \pm5.4115.13±25.8715.13 \pm25.872.7±3.22.7 \pm3.2
Juniper0.000.0000.000.000
Gurobi AC-OTS0.000.0000.000.000
Gurobi DC-OTS0.000.0000.000.000
Reference (24h)
Juniper0.00‡0.00^{\ddagger}0.0000.00‡0.00^{\ddagger}0.000

Table 2. Generation-cost savings, load-payment savings, and number of opened lines for the Texas2k–data centers scenario and the Texas2k–Fort Worth halved scenarios, with a 50 s time limit. Relative to the nominal summer-peak case (85,758.9 MW), the Texas2k–data centers scenario and the Texas2k–Fort Worth halved scenarios increases the load by 10.7% and 9.9%, respectively. Seeded baselines report mean ±\pm standard deviation over 10 seeds. ‡: no integral incumbent within 24 h.

Texas2k data-center discussion

Under the Texas2k–data center scenario, the load payment is $24.34M per hour for the base topology. Under the topology obtained with SES it pays $1.38M per hour, a 94.34% reduction. Figure 4 shows nodal prices of the intact topology, topology found by the strongest baseline (PSeq), and topology found by our method (SES). With the intact topology, prices exceed $100/MWh at3 580 of the 2,751 buses, exceed $1,000/MWh at 44 buses, and peaks at $13,316/MWh. The area with higher nodal prices forms a corridor that runs from Amarillo through Lubbock and Abilene into Dallas–Fort Worth and south toward Waco, with the most expensive buses concentrated around Abilene and Dallas–Fort Worth. With the five openings obtained by SES within 50 s, the corridor with high prices clears: with only 2 buses remain above $100/MWh, and the maximum drops to $169/MWh. The topology obtained by PSeq reduces the prices, but prices remain high at the Dallas–Fort Worth and Abilene areas. The corresponding load payment at each bus is shown in Figure 5.

Nodal prices on the Texas2k-data centers scenario

Figure 4. Nodal prices on the Texas2k-data centers scenario: the intact network (left), after switching with the PSeq baseline (1 line open, middle) and with SES (ours, 5 lines open, right), each at 50 s. Each dot is a bus colored by its nodal price. The color bands are shared by the three panels and placed at quantiles of their pooled prices, rounded to readable values (edges at $−80, $−30, 0, 10, 12, 15, 20, 25, 30, 40, 50, 70 and 150/MWh), so each color covers a similar number of buses; negative prices are blue, positive prices run from teal through yellow to dark red, and the color bar draws every band at equal width. In the intact topology, 580 of the 2,751 buses exceed $100/MWh (maximum $13,316/MWh) and the median price is $32.91/MWh; with PSeq the median falls to $22.14/MWh but 56 buses around Dallas–Fort Worth and Abilene stay above $100/MWh; with SES the median is $12.46/MWh and only 2 buses exceed $100/MWh.

Load payment per bus on the Texas2k-data centers scenario

Figure 5. Load payment per bus on the Texas2k-data centers scenario: the intact network (left), after switching with the PSeq baseline (1 line open, middle) and with SES (ours, 5 lines open, right), each at 50 s. Each dot is a bus with load, colored by its load payment per hour, defined as the bus’s nodal price multiplied by the load served at that bus; the total load payment is the sum over all buses. Color bands are placed at quantiles of the pooled payments of the three panels. Total load payment is $24.34M/h intact, $5.33M/h with PSeq and $1.38M/h with SES; buses paying more than $100k/h fall from 55 to 5 to none.

Figure 6 maps the change in load payment at each bus (blue: pays less; red: pays more; circle area proportional to the change in $/h). In the data centers scenario, load pays less at 1,083 buses and more at 129. The reductions are concentrated where nodal prices were highest: 94 % of the gross payment reduction falls at the 311 buses whose intact price exceeded $100/MWh, roughly 64 % of it in Dallas–Fort Worth and 29 % around Abilene. The two largest single reductions, about $2.9M/h and $2.8M/h, are at Abilene. At the same time, a small number of buses pay modestly more (buses east of Dallas toward Tyler and some near the Abilene hub). The buses that pay more are almost all ones with a negative nodal price on the intact network (119 of the 129).

Change in load payment per bus after switching on Texas2k-data centers

Figure 6. Change in load payment per bus after switching on Texas2k-data centers (SES, 5 lines open, 50 s time budget). Each circle is a bus with load; its area is proportional to the change in what that load pays per hour (nodal price ×\times load), relative to the intact network. Blue: the bus pays less after switching (1,083 buses); red: it pays more (129 buses). Total load payment falls from $24.34M/h to $1.38M/h.

Conclusion and Discussion

We develop a novel learning-guided method to solve the AC Optimal Transmission Switching problem within real-time computational budgets. Our fine-tuning pipeline adapts a pretrained AC-OPF surrogate into a grid-specific switching-value model (Line-switching Ranking Proxy) that estimates the impact of topology changes on generation cost. Domain-informed filtering keeps ML inference tractable, while fast ML-based evaluation enables exploration of a broader set of candidates. didate lines than conventional methods. Within this search space, learning-guided local search prioritizes promising switching moves, which are subsequently verified using full AC-OPF solves to ensure feasibility and improvements in generation cost. Experiments on existing benchmarks and realistic load scenarios of the 2,000-bus Texas network show that the proposed ML-guided framework identifies switching decisions that significantly reduce the generation cost within a 50-second computational budget. Our results illustrate how ML and optimization can be combined to explore large-scale combinatorial search in power systems, offering a promising path toward tackling increasingly complex grid-control and planning problems at operationally relevant scales.

We also acknowledge several limitations. First, all benchmark networks considered in this study are synthetic test systems. Second, the search does not enforce (NN-1) security of the resulting switched topologies [65]. Third, while every accepted topology is verified to be AC-feasible, the proposed methods provide no guarantees of global optimality. Finally, we evaluate switched topologies as static operating points; transient stability, protection coordination, and the operational sequencing of multiple switching actions are outside the scope of the current model. Extending the framework to security-constrained switching, dynamic-security considerations, and multi-period operation is a natural direction for future work.

Appendix

Texas2k case construction

Texas2k–data centers

Data center growth differs from conventional demand growth in both its spatial concentration and the speed at which large increments of load can appear. Individual facilities and campuses can reach hundreds of megawatts or more, potentially creating new transmission bottlenecks or substantially altering existing congestion patterns. Although Texas2k is a synthetic network, it is designed to approximate the structure of the ERCOT transmission system and provides geographic locations for buses and transmission lines [55]. We therefore use it to construct a geographically grounded data center stress case based on publicly announced projects available as of September 2026.

At the time of writing, data center-related large-load interconnection requests in the Electric Reliability Council of Texas (ERCOT) service area exceed 420 GW[66]. This is several times larger than ERCOT’s approximately 91 GW system peak demand, and the raw interconnection queue should therefore not be interpreted as a forecast of realized load. Rather than modeling the full queue, we use publicly announced projects for which both location and planned capacity are available[57, 58, 59, 60].

We distribute the resulting build-out across ten geographic hubs according to the announced project pipeline where project-level information is available and according to established data center market shares elsewhere. This construction is intended to represent the geography of data center development rather than population. For example, Abilene receives the same 20% share as Dallas despite its substantially smaller population because of its announced campus pipeline: approximately 1.2 GW has been developed, with a further 0.9 GW announced toward a total of 2.1 GW[57, 58]. The remaining hub shares, their project-level basis, and the source snapshot used to construct them are documented in the released case-construction script and accompanying README.

Within each hub, new demand is assigned to geographically corresponding Texas2k buses using the location labels supplied with the dataset. We use load-serving buses at 115 and 161 kV and any geographically matching bus at 230 and 500 kV. Individual injections are capped according to interconnection voltage, from 300 MW at 115 kV to 2,000 MW at 500 kV. These caps are scenario assumptions intended to prevent unrealistically large injections at lower-voltage transmission buses; very large campuses would in practice connect through higher-voltage infrastructure rather than through a single low-voltage substation. Each campus is represented as a lumped active and reactive power injection at the transmission bus with a power factor of 0.98. Customer substations, step-down transformers, collector networks and behind-the-meter generation are not modelled. The underlying Texas2k generation fleet and transmission network are otherwise unchanged.

We scale the total added data center demand by bisection to identify a highly stressed but AC-feasible operating point. Starting from the Texas2k 2025 summer-peak case, the total data center build-out is increased until the intact-network AC-OPF no longer returns a locally converged feasible operating point. The largest converged build-out is 9,139.5 MW, corresponding to a 10.7% increase relative to the original 85.8 GW summer-peak demand. We use this point as the Texas2k–data centers scenario. The resulting total active demand is 94,898.4 MW, and the geographic distribution of the added load is shown in Figure 7.

Constructed 9.1-GW data center stress scenario

Figure 7. Constructed 9.1-GW data center stress scenario. 149 synthetic load injections across ten metro clusters, colored by interconnection voltage and sized by MW, with the allocation by hub. The allocation follows announced projects and market weight, not population; individual points are eligible buses of the synthetic grid, not announced project locations.

Texas2k–Fort Worth halved

The Texas2k–data centers scenario represents a single highly stressed operating point near the network’s estimated hosting-capacity boundary. We therefore evaluate sensitivity to uncertainty in the magnitude of geographically concentrated demand. The Dallas–Fort Worth metroplex is the dominant data center market in Texas and one of the largest in the United States [67], but announced data center capacity is highly uncertain and may not fully materialize. To represent a lower realization of the announced Fort Worth build-out, we construct the Texas2k–Fort Worth halved scenario by reducing all added data center load assigned to the Fort Worth hub to 50% of its value in the Texas2k–data centers scenario. Added demand at all other hubs is left unchanged. This perturbation isolates the sensitivity of the switching solutions to uncertainty in one of the largest geographically concentrated load centers.

Pseudocode for GridSFM-guided methods

Notation. Let E′⊆EE' \subseteq E denote the set of candidate lines described in § 3.2, excluding lines whose removal islands the intact network. Let S⊆E′S \subseteq E' denote the set of currently opened lines, JJ the maximum number of simultaneously opened lines, and TT the wall-clock time budget. Given the current open set SS, let I(S)\mathcal{I}(S) denote the set of additional lines whose removal would island the network.

We use C^θ(S)\widehat{C}_{\theta}(S) to denote the GridSFM-predicted AC-OPF cost of the topology with lines SS open. Predictions for multiple candidate topologies are evaluated in a single batch. We use C(S)C(S) to denote the cost returned by an exact AC-OPF solve and set C(S)=+∞C(S) = +\infty if the solve is infeasible. or fails to converge. The current best verified AC-OPF cost found so far is denoted by c⋆c^\star. A candidate topology is accepted only if C(S)<c⋆C(S) < c^\star.

The pseudocode for Fixed-Rank Search and Adaptive-Rank Search are shown in Algorithm 1 and Algorithm 2, respectively. Swap-Enhanced Search is shown in Algorithm 3, and it includes a Fallback procedure. In Swap-Enhanced Search, if none of the shortlisted moves improves the incumbent, Fallback scans the full fixed ranking π\pi, skipping already opened lines and previously solved or islanding topologies. When ∣S∣<J|S| < J, it tests pure line openings; when ∣S∣=J|S| = J, it tests swaps between each candidate and the AA most recent openings in QQ. It returns the first topology with verified cost below c⋆c^\star, or ∅\varnothing if none is found before the ranking or time budget is exhausted.

Algorithm 1 Fixed-Rank Search (FRS)
Require: candidate set $\mathcal{E}$, line limit $J$, time budget $T$

1: $S \leftarrow \varnothing$,    $c^\star \leftarrow C(\varnothing)$
2: $\pi \leftarrow \operatorname{Sort}(\mathcal{E}, b \mapsto \widehat{C}_{\theta}(\{b\}))$
    ▷ ascending predicted cost; one-time ranking
3: for $b$ in order $\pi$ do
4:     if $|S| = J$ or time $\ge T$ then
5:         break
6:     end if
7:     if $b \in \mathcal{I}(S)$ then
8:         continue
9:     end if
10:    $c \leftarrow C(S \cup \{b\})$    ▷ solver verification
11:    if $c < c^\star$ then
12:        $S \leftarrow S \cup \{b\}$,    $c^\star \leftarrow c$
13:    end if
14: end for
15: return $S$
Algorithm 2 Adaptive-Rank Search (ARS)
Require: candidate set $\mathcal{E}$, line limit $J$, time budget $T$
1: $S \leftarrow \varnothing$, $c^\star \leftarrow C(\varnothing)$
2: while $|S| < J$ and time $< T$ do
3: $\pi \leftarrow \operatorname{Sort}(\mathcal{E} \setminus (S \cup \mathcal{I}(S)), b \mapsto \widehat{C}_{\theta}(S \cup \{b\}))$ ▷ ascending predicted cost, rerank
4: if $\pi = \varnothing$ then
5:     break
6: end if
7: $\mathit{improved} \leftarrow \mathit{false}$
8: for $b$ in order $\pi$ do
9:     if time $\ge T$ then
10:         break
11:     end if
12:     $c \leftarrow C(S \cup \{b\})$ ▷ solver verification
13:     if $c < c^\star$ then
14:         $S \leftarrow S \cup \{b\}$, $c^\star \leftarrow c$
15:         $\mathit{improved} \leftarrow \mathit{true}$
16:         break ▷ accept first improvement, then rerank
17:     end if
18: end for
19: if not $\mathit{improved}$ then
20:     break ▷ no improving line removal found
21: end if
22: end while
23: return $S$

Algorithm 3 Swap-Enhanced Search (SES)

Require: candidate set E\mathcal{E}, line limit JJ, time budget TT; Amax⁡=5A_{\max}=5, M=10M=10, L=2L=2

  1. A←min⁡(Amax⁡,J)A \leftarrow\min(A_{\max}, J)

  2. S←∅S \leftarrow\varnothing, c⋆←C(∅)c^\star\leftarrow C(\varnothing)

  3. π←Sort⁡(E,b↦C^θ({b}))\pi\leftarrow\operatorname{Sort}(\mathcal{E}, b \mapsto\widehat{C}_{\theta}(\{b\})) ▷ fixed ranking

Phase 1: greedy

  1. for bb in order π\pi do

  2. if ∣S∣=A|S|=A or time ≥T\ge T then break

  3. end if

  4. if b∈I(S)b \in\mathcal{I}(S) then continue

  5. end if

  6. c←C(S∪{b})c \leftarrow C(S \cup\{b\})

  7. if c<c⋆c<c^\star then

  8. S←S∪{b}S \leftarrow S \cup\{b\}, c⋆←cc^\star\leftarrow c

  9. end if

  10. end for

Phase 2: open/swap local search

  1. while time <T<T do

  2. B←B \leftarrow first MM lines in π\pi not in SS

  3. Q←Q \leftarrow up to AA most recently opened lines in SS

  4. Nopen←{S∪{b}:b∈B, ∣S∣<J, b∉I(S)}\mathcal{N}_{\mathrm{open}} \leftarrow\{S \cup\{b\}: b \in B,\ |S|<J,\ b \notin\mathcal{I}(S)\}

  5. Nswap←{(S∖{a})∪{b}:a∈Q, b∈B, b∉I(S∖{a})}\mathcal{N}_{\mathrm{swap}} \leftarrow\{(S \setminus\{a\}) \cup\{b\}: a \in Q,\ b \in B,\ b \notin\mathcal{I}(S \setminus\{a\})\}

  6. L←Top⁡L(Nopen,C^θ)∪Top⁡L(Nswap,C^θ)\mathcal{L} \leftarrow\operatorname{Top}_{L}(\mathcal{N}_{\mathrm{open}},\widehat{C}_{\theta}) \cup\operatorname{Top}_{L}(\mathcal{N}_{\mathrm{swap}},\widehat{C}_{\theta}) ▷ surrogate shortlist

  7. exactly solve all previously unsolved S′∈LS' \in\mathcal{L}

  8. if min⁡S′∈LC(S′)<c⋆\min_{S' \in\mathcal{L}} C(S')<c^\star then

  9. S←arg⁡min⁡S′∈LC(S′)S \leftarrow\arg\min_{S' \in\mathcal{L}} C(S'), c⋆←C(S)c^\star\leftarrow C(S)

  10. else

  11. S′←Fallback⁡(S,π,Q,J,T)S' \leftarrow\operatorname{Fallback}(S,\pi,Q,J,T) ▷ first verified improvement

  12. if S′=∅S'=\varnothing then

  13. break

  14. end if

  15. S←S′S \leftarrow S', c⋆←C(S′)c^\star\leftarrow C(S')

  16. end if

  17. end while

  18. return SS

Data availability

All two Texas2k evaluation cases, their generation receipts, the runs behind every table entry and figure, and complete per-grid recipes that regenerate the training data will be released with the code upon publication, with an archived private snapshot available to reviewers. Benchmark networks: PGLib-OPF (https://github.com/power-grid-lib/pglib-opf) and the Texas A&M ACTIVSg2000/Texas2k repository (https://electricgrids.engr.tamu.edu/electric-grid-test-cases/activsg2000/).

Code availability

The code for the search methods, the GridSFM fine-tuning recipe, the AC-OPF verification pipeline, all baselines and the case-construction scripts is available to reviewers through an archived private snapshot and will be publicly released upon publication.

Acknowledgments

This work was done during Weimin Huang’s internship at Microsoft Research and at the University of Southern California. The research is partially supported by the National Science Foundation (NSF) grant #2112533: “NSF Artificial Intelligence Research Institute for Advances in Optimization”, the NSF grant #2346058: “NRT-AI: Integrating Artificial Intelligence and Operations Research Technologies”.

Author contributions

Weimin Huang developed the learning-guided search methods, conducted the experiments and co-wrote the manuscript.

Weiwei Yang adapted and fine-tuned GridSFM to construct the Line-switching Ranking Proxy, set up the initial experiments and datasets, and co-wrote the manuscript.

Bistra Dilkina provided research oversight, gave feedback on the optimization methodology and edited the manuscript.

Konstantina Mellou mentored Weimin Huang, advised on the design of the search methods and edited the manuscript.

Ishai Menache provided research oversight, gave feedback on the optimization methodology and edited the manuscript.

Marco Molinaro mentored Weimin Huang, advised on the design of the search methods, assisted with the experiments and edited the manuscript.

Chris White provided research oversight and edited the manuscript.

Baosen Zhang provided research oversight and power systems and AI expertise, and co-wrote the manuscript.

Competing interests

The authors declare no competing interests.

Correspondence

Correspondence and requests for materials should be addressed to Weiwei Yang ([email protected]).

References

  1. [1]Eric O’Shaughnessy, Ryan H. Wiser, Peter Cappers, Jeffrey Deason, Natalie Mims Frick, and Ryan Hledik. Revisiting the relationship between demand growth and electricity prices. Joule, 10(7):102528, July 2026. doi: 10.1016/j.joule.2026.102528. URL https://doi.org/10.1016/j.joule.2026.102528.
  2. [2]Anton Tausendfreund, Sarah Schreyer, Florian Immig, Ulrich Oberhofer, Julius Trebbien, Aaron Praktiknjo, Benjamin Schäfer, and Dirk Witthaut. Understanding the european energy crisis through structural causal models. arXiv preprint arXiv:2506.00680, 2025. doi: 10.48550/arXiv.2506.00680. URL https://arxiv.org/abs/2506.00680.
  3. [3]Laila Kearney. Prices jump 22% in biggest US power grid energy auction, July 2025. URL https://www.reuters.com/business/energy/prices-jump-22-biggest-us-power-grid-energy-auction-2025-07-22/. Reuters.
  4. [4]International Energy Agency. Electricity 2026: Analysis and forecast to 2030. Technical report, International Energy Agency, Paris, February 2026. URL https://www.iea.org/reports/electricity-2026.
  5. [5]June Lukuyu and Jay Taneja. Powering up in africa: electricity demand stimulation to build inclusive economies, reduce inequality, and improve power system sustainability. Current Sustainable/renewable Energy Reports, 10(1):1–11, 2023.DOI
  6. [6]Ryan H. Wiser, Eric O’Shaughnessy, Galen L. Barbose, Peter Cappers, and Will Gorman. Factors influencing recent trends in retail electricity prices in the united states. The Electricity Journal, 38(4):107516, 2025. doi: 10.1016/j.tej.2025.107516. URL https://doi.org/10.1016/j.tej.2025.107516.
  7. [7]European Union Agency for the Cooperation of Energy Regulators. Key developments in the EU energy sector: What are the options to further reduce energy prices? Presentation to the transport, telecommunications and energy council ministerial, European Union Transport, Telecommunications and Energy Council, Ljubljana, Slovenia, March 2026. URL https://www.acer.europa.eu/sites/default/files/documents/en/The_agency/Documents/20260316-ACER-Presentation-TTE-Council-Key-developments.pdf.
  8. [8]Federal Energy Regulatory Commission. 2025 state of the markets. Technical report, Federal Energy Regulatory Commission, March 2026. URL https://www.ferc.gov/sites/default/files/2026-03/26_State-of-the-Market_0324_1430.pdf. See Figure 5, p. 7, for in-service data center capacity.
  9. [9]Kimberly Peterson. U.S. electricity generation in 2025 hit a record, again. U.S. Energy Information Administration, Today in Energy, March 2026. URL https://www.eia.gov/todayinenergy/detail.php?id=67284. Published March 5, 2026.
  10. [10]U.S. Energy Information Administration. EIA open data, 2026. URL https://www.eia.gov/opendata/. Accessed: September 28, 2026.
  11. [11]U.S. Department of Energy. Transforming the nation’s electricity system: The second installment of the quadrennial energy review. Technical report, U.S. Department of Energy, Washington, DC, January 2017. URL https://www.energy.gov/policy/quadrennial-energy-review-second-installment.
  12. [12]Emilia Chojkiewicz, Umed Paliwal, Nikit Abhyankar, Casey Baker, Ric O’Connell, Duncan Callaway, and Amol Phadke. Accelerating transmission capacity expansion by using advanced conductors in existing right-of-way. Proceedings of the National Academy of Sciences, 121(40):e2411207121, 2024. doi: 10.1073/pnas.2411207121.DOI
  13. [13]Daniel Shawhan, McKenna Peplinski, Sally Robson, Ethan Russell, Ethan Ziegler, Maya Domeshek, and Karen Palmer. Clean power delayed: Effects of infrastructure delays on health, environment, and US households. Working paper, Resources for the Future, May 2025. URL https://www.rff.org/publications/working-papers/clean-power-delayed-effects-of-infrastructure-delays-on-health-environment-and-us-households/.
  14. [14]Emily B Fisher, Richard P O’Neill, and Michael C Ferris. Optimal transmission switching. IEEE Transactions on Power Systems, 23(3):1346–1355, 2008.
  15. [15]J David Fuller, Raynier Ramasra, and Amanda Cha. Fast heuristics for transmission-line switching. IEEE Transactions on Power Systems, 27(3):1377–1386, 2012.DOI
  16. [16]Weiwei Yang, Andrea Britto Mattos Lima, Thiago V. Spina, Spencer Fowers, Baosen Zhang, and Chris White. GridSFM: A foundation model for AC optimal power flow. Microsoft Research technical report, May 2026. URL https://www.microsoft.com/en-us/research/publication/gridsfm-a-foundation-model-for-ac-optimal-power-flow/.
  17. [17]Luke Bhan, Weiwei Yang, Margaret Capetz, and Baosen Zhang. Gridsfm: A foundation model for solving ac optimal power flow, 2026. URL https://arxiv.org/abs/2609.30173.
  18. [18]J Duncan Glover, Mulukutla S Sarma, Thomas Jeffrey Overbye, and NP Padhy. Power system analysis and design, volume 2008. Cengage Learning Stamford, CT, USA, 2012.
  19. [19]Weimin Huang, Taoan Huang, Aaron M Ferber, and Bistra Dilkina. Distributional MIPLIB: A multi-domain library for advancing ML-guided MILP methods. arXiv:2406.06954, 2024.arxiv.org/abs/2406.06954
  20. [20]Lara Scavuzzo, Karen Aardal, Andrea Lodi, and Neil Yorke-Smith. Machine learning augmented branch and bound for mixed integer linear programming. Mathematical Programming, pages 1–44, 2024.
  21. [21]Yoshua Bengio, Andrea Lodi, and Antoine Prouvost. Machine learning for combinatorial optimization: a methodological tour d’horizon. European Journal of Operational Research, 290(2):405–421, 2021.
  22. [22]Weimin Huang, Ryan Piansky, Bistra Dilkina, and Daniel K Molzahn. Machine learning guided optimal transmission switching to mitigate wildfire ignition risk. Electric Power Systems Research, 262:113687, 2027.
  23. [23]Al-Amin B Bugaje, Jochen L Cremer, and Goran Strbac. Real-time transmission switching with neural networks. IET Generation, Transmission & Distribution, 17(3):696–705, 2023.
  24. [24]Tong Han and David J Hill. Learning-based topology optimization of power networks. IEEE Transactions on Power Systems, 38(2):1366–1378, 2023. URL https://doi.org/10.1109/TPWRS.2022.3170083.
  25. [25]Vincenzo Di Vito, Mehdi Taghizadeh, Deepjyoti Deka, Kaarthik Sundar, and Ferdinando Fioretto. Learning discrete decisions for MIPs with constraint-aware diffusion. arXiv preprint arXiv:2608.13079, 2026.
  26. [26]Kyri Baker. Emulating ac opf solvers for obtaining sub-second feasible, near-optimal solutions. arXiv preprint arXiv:2012.10031, 2020.arxiv.org/abs/2012.10031
  27. [27]Burak Kocuk, Santanu S Dey, and Xu Andy Sun. New formulation and strong MISOCP relaxations for AC optimal transmission switching problem. IEEE Transactions on Power Systems, 32(6):4161–4170, 2017.DOI
  28. [28]Pranav Jain, Peixuan Zhang, Devon Sigler, and Bernard Knueven. Acopf transmission switching using open-source minlp solvers. In 2024 IEEE Power & Energy Society General Meeting (PESGM), pages 1–5. IEEE, 2024.DOI
  29. [29]Xin Chen, Xiaoyang Wang, Ana Colacelli, Matt Lee, and Le Xie. Electricity demand and grid impacts of ai data centers: Challenges and prospects. arXiv preprint arXiv:2509.07218, 2025.arxiv.org/abs/2509.07218
  30. [30]Ruimin Zhu, Keyu Zhai, Ning Lin, and Venkata Dinavahi. Evolving from heavy power burden to grid resilience supporter: A review of the role of data centers interacting with power grid. Renewable and Sustainable Energy Reviews, 242:117298, 2026.
  31. [31]Hassan Hijazi and Sylvie Thiébaux. Optimal distribution systems reconfiguration for radial and meshed grids. International Journal of Electrical Power & Energy Systems, 72:136–143, 2015.DOI
  32. [32]Soumya Das, Ashu Verma, and PR Bijwe. Security constrained ac transmission network expansion planning. Electric Power Systems Research, 172:277–289, 2019.DOI
  33. [33]Kory W Hedman, Shmuel S Oren, and Richard P O’Neill. A review of transmission switching and network topology optimization. In 2011 IEEE Power and Energy Society General Meeting, pages 1–7. IEEE, 2011.DOI
  34. [34]Steven Low. Power System Analysis - Analytical Tools and Structural Properties. Cambridge University Press, 2026.
  35. [35]Florin Capitanescu. Critical review of recent advances and further developments needed in ac optimal power flow. Electric Power Systems Research, 136:57–68, 2016.
  36. [36]Carleton Coffrin, Hassan L Hijazi, and Pascal Van Hentenryck. The QC relaxation: A theoretical and computational study on optimal power flow. IEEE Transactions on Power Systems, 31(4):3008–3018, 2016.arxiv.org/abs/1502.07847
  37. [37]Mary B Cain, Richard P O’Neill, and Anya Castillo. History of optimal power flow and formulations. Staff technical paper, Federal Energy Regulatory Commission, 2012.
  38. [38]Javad Lavaei and Steven H Low. Zero duality gap in optimal power flow problem. IEEE Transactions on Power systems, 27(1):92–107, 2011.DOI
  39. [39]Line A Roald, David Pozo, Anthony Papavasiliou, Daniel K Molzahn, Jalal Kazempour, and Antonio Conejo. Power systems optimization under uncertainty: A review of methods and applications. Electric Power Systems Research, 214:108725, 2023.DOI
  40. [40]IEEE Power and Energy Society. IEEE Standard for Ratings and Requirements for AC High-Voltage Circuit Breakers with Rated Maximum Voltage Above 1000 V. Technical Report IEEE Std C37.04-2018, New York, NY, USA, May 2019. URL https://standards.ieee.org/ieee/C37.04/5357/.
  41. [41]Jesse T. Holzer, Stephen Elbert, Hans Mittelmann, Richard O’Neill, and HyungSeon Oh. GO competition challenge 3: problem, solvers, and solution analysis. Energy Systems, January 2025. doi: 10.1007/s12667-024-00708-1. URL https://doi.org/10.1007/s12667-024-00708-1.
  42. [42]Samuel Chevalier. Low-rank and lifted semidefinite programming for mixed-integer polynomial power grid optimization, September 2026. URL https://arxiv.org/abs/2609.38435.
  43. [43]Zhu Yang and Shmuel Oren. Line selection and algorithm selection for transmission switching by machine learning methods. In 2019 IEEE Milan PowerTech, 2019.DOI
  44. [44]Minsoo Kim et al. Dispatch-aware learning for optimal transmission switching. arXiv preprint arXiv:2512.17516, 2025.arxiv.org/abs/2512.17516
  45. [45]Mostafa Soroush and J David Fuller. Accuracies of optimal transmission switching heuristics based on DCOPF and ACOPF. IEEE Transactions on Power Systems, 29(2):924–932, 2014.DOI
  46. [46]Babak Taheri and Daniel K Molzahn. AC-informed DC optimal transmission switching problems via parameter optimization. IEEE Transactions on Power Systems, 40(6):5422–5433, 2025.
  47. [47]Rahul K Gupta. AC-informed DC optimal transmission switching via admittance sensitivity-augmented constraints and repair costs. arXiv preprint arXiv:2603.01249, 2026.arxiv.org/abs/2603.01249
  48. [48]Xiang Pan, Minghua Chen, Tianyu Zhao, and Steven H Low. Deepopf: A feasibility-optimized deep neural network approach for ac optimal power flow problems. IEEE Systems Journal, 17 (1):673–683, 2022.
  49. [49]Min Zhou, Minghua Chen, and Steven H Low. Deepopf-ft: One deep neural network for multiple ac-opf problems with flexible topology. IEEE Transactions on Power Systems, 38(1): 964–967, 2022.
  50. [50]Mostafa Sahraei-Ardakani, Xingpeng Li, P Balasubramanian, KW Hedman, and Mojdeh Abdi-Khorsand. Real-time contingency analysis with transmission switching on real power system data. IEEE Transactions on Power Systems, 31(3):2501–2502, 2015.DOI
  51. [51]Xingpeng Li, Pranavamoorthy Balasubramanian, Mostafa Sahraei-Ardakani, Mojdeh Abdi-Khorsand, Kory W Hedman, and Robin Podmore. Real-time contingency analysis with corrective transmission switching. IEEE Transactions on Power Systems, 32(4):2604–2617, 2016.
  52. [52]Jiachun Guo, Yong Fu, Zuyi Li, and Mohammad Shahidehpour. Direct calculation of line outage distribution factors. IEEE Transactions on Power Systems, 24(3):1633–1634, 2009.DOI
  53. [53]Sogol Babaeinejad sarookolaee, Byungkwon Park, Bernard Lesieutre, and Christopher L DeMarco. Transmission congestion management via node-breaker topology control. IEEE Systems Journal, 17(3):3413–3424, 2023.DOI
  54. [54]Sogol Babaeinejad sarookolaee, Adam Birchfield, Richard D Christie, Carlton Coffrin, Christopher DeMarco, Ruisheng Diao, Michael Ferris, Stephane Fliscounakis, Scott Greene, Renke Huang, et al. The power grid library for benchmarking AC optimal power flow algorithms. arXiv preprint arXiv:1908.02788, 2019. URL https://arxiv.org/abs/1908.02788. Benchmark library at https://github.com/power-grid-lib/pglib-opf.
  55. [55]Adam B Birchfield, Ti Xu, Kathleen M Gegner, Komal S Shetye, and Thomas J Overbye. Grid structural characteristics as validation criteria for synthetic networks. IEEE Transactions on Power Systems, 32(4):3258–3265, 2017.DOI
  56. [56]Ti Xu, Adam B Birchfield, Kathleen M Gegner, Komal S Shetye, and Thomas J Overbye. Application of large-scale synthetic power system models for energy economic studies. In Proceedings of the 50th Hawaii International Conference on System Sciences (HICSS), 2017.
  57. [57]Crusoe Energy Systems. Crusoe announces new 900 MW AI factory campus in abilene, texas to support microsoft AI infrastructure. Press release, March 2026. URL https://www.crusoe.ai/resources/newsroom/crusoe-announces-new-900-mw-ai-factory-campus-in-abilene-texas-to-support-microsoft-ai-infrastructure.
  58. [58]DataCenterDynamics. Crusoe tops out final building at OpenAI Stargate data center campus in abilene, texas. DataCenterDynamics, November 2025. URL https://www.datacenterdynamics.com/en/news/crusoe-tops-out-final-building-at-openai-stargate-data-center-campus-in-abilene-texas/.
  59. [59]Energy Capital Partners and KKR. Energy capital partners and KKR announce development of hyperscale data center campus in texas. Press release, Business Wire, July 2025. URL https://www.morningstar.com/news/business-wire/20250730057911/energy-capital-partners-and-kkr-announce-development-of-hyperscale-data-center-campus-in-texas.
  60. [60]Stream Data Centers. Stream data centers breaks ground on new hyperscale data center campus in san antonio. Press release, June 2024. URL https://www.streamdatacenters.com/news/breaks-ground-hyperscale-campus-san-antonio/.
  61. [61]Ole Kröger, Carleton Coffrin, Hassan Hijazi, and Harsha Nagarajan. Juniper: An open-source nonlinear branch-and-bound solver in julia. In International conference on the integration of constraint programming, artificial intelligence, and operations research, pages 377–386. Springer, 2018.DOI
  62. [62]Carleton Coffrin, Russell Bent, Kaarthik Sundar, Yee sian Ng, and Miles Lubin. Powermodels.jl: An open-source framework for exploring power flow formulations. In 2018 Power Systems Computation Conference (PSCC), pages 1–8. IEEE, 2018.
  63. [63]Daniel S Kirschen and Goran Strbac. Fundamentals of power system economics. John Wiley & Sons, 2026.
  64. [64]Sungho Shin, Mihai Anitescu, and François Pacaud. Accelerating optimal power flow with GPUs: SIMD abstraction of nonlinear programs and condensed-space interior-point methods. Electric Power Systems Research, 236:110651, 2024. URL https://doi.org/10.1016/j.epsr.2024.110651.
  65. [65]Mojtaba Khanabadi, Hassan Ghasemi, and Meysam Doostizadeh. Optimal transmission switching considering voltage security and N-1 contingency analysis. IEEE Transactions on Power Systems, 28(1):542–550, 2013.DOI
  66. [66]ERCOT. Assessing the grid: ERCOT update to the Texas Senate, 2026. URL https://www.ercot.com/files/docs/2026/07/29/ERCOT-Senate-July-29-Panel-1-Assessing-The-Grid.pdf. July 2026 legislative update; approximately 474 GW of large-load interconnection requests, about 90% associated with datacenters.
  67. [67]CBRE. Global data center trends 2026. Technical report, CBRE, 2026.

Paper details

vlnc:W_4ruse6d4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents