When engineers want a service to survive a failure, they give it two routes that share no cables or switches. In a power grid, that promise does not hold. Electricity does not follow the route an operator draws. It spreads over every available loop according to Kirchhoff's laws, so when one line trips, its flow jumps onto other lines, sometimes far away. Two supply routes to a hospital or a data center can share no hardware and still fail together: a single outage elsewhere can push both of them past their limits, or the trip of one can overload the other. This note proposes a design problem and a solver for it. The problem is to choose a small set of transmission lines to switch off, together with $k$ physically separate supply corridors, so that no single outage, including the first round of overload trips it causes, brings down all the corridors at once. The solver scales because the power-systems theory of failure localization (which lines an outage can and cannot disturb) lets us discard most outages exactly, before any optimization runs.
The opportunity comes from a structural fact that sits between two literatures. Work on failure localization proves that, under the standard linearized ("DC") power flow model, an outage on line $f$ changes the flow on line $e$ if and only if some simple cycle contains both lines (Line Failure Localization of Power Networks Part I: Non-Cut Outages). Two internally disjoint paths from a source $s$ to a sink $t$ together form a cycle. Disjoint corridors are therefore always electrically coupled, and ordinary disjoint-path routing cannot give flow independence. Independence has to be engineered: by changing the topology, and by bounding the coupling quantitatively rather than asking for none.
Core Claim
We claim that cascade-aware corridor design can be written as a mixed-integer program (an optimization problem with both yes/no and continuous decisions) that people can actually solve. The program chooses line openings and corridor paths subject to post-outage capacity constraints. Two exact tools from the localization theory make it tractable. The first is a block decomposition, which splits the grid into biconnected pieces that stay connected after removing any one node. It tells us which outages can touch corridor lines at all. The second is a closed-form sensitivity of each line to each outage, the Line Outage Distribution Factor (LODF). It turns "does outage $f$ overload corridor line $e$" into a linear constraint. Against plain node-disjoint routing, we expect a large gain in the fraction of single outages after which at least one corridor still delivers its transfer. We also expect solve times that grow with the size of the corridor-touching blocks, not with the size of the whole grid.
Relation to Prior Work
The localization line of work motivates the proposal and supplies its main method. Less is More: Real-time Failure Localization in Power Systems opens tie-lines to turn the area structure of a meshed grid into a tree, so that disturbances stay inside the area where they start. Adaptive Network Response to Line Failures in Power Systems aligns balancing areas with the grid's biconnected blocks and adds a fast controller that guarantees localization for non-critical failures. Both papers optimize for containment, and they pay for it with redundancy: a tree has single points of failure, which the file lists as an open limitation. Our objective runs the other way. We keep deliberate redundancy for a protected transfer and bound the coupling it creates.
Line Failure Localization of Power Networks Part II: Cut Set Outages extends the analysis to outages that split the grid into islands. Its simple-path criterion gives the exact screening rule we need for those outages.
Interface Networks for Failure Localization in Power Systems shows that a 2×2 bipartite interface between two sub-grids can have zero net cross-grid LODF even though it keeps parallel paths, provided the line susceptances (the electrical ease of flow along each line) satisfy a rank-1 condition. This is the one known construction in which redundant paths are electrically decoupled. We use it as a design primitive: at an area boundary where the two corridors cross, the solver may impose the rank-1 match instead of opening a line.
The disjoint-routing literature constrains the design. Work on path uniqueness (Chen et al.) and on node- versus edge-disjoint protection (Yang et al.; Aly et al.) measures independence purely by shared graph elements. That is exactly the blind spot the cycle argument above exposes. ShareDP makes batch $k$-disjoint path queries fast by sharing traversals over a merged split graph (a copy of the network with each node doubled to enforce node-disjointness). It works on static, uncapacitated graphs, so it serves as a candidate generator here, not as a solution.
The scalability literature supplies the optimization style. Optimal QoS-Aware Network Slicing for Service-Oriented Networks with Flexible Routing shows that exact reformulations plus valid inequalities (extra constraints that tighten the relaxation without removing any integer solution) can make joint placement and routing tractable in Gurobi. We follow that approach of exact reformulation rather than heuristics.
Research Program
The model uses the DC power flow on a graph $G=(V,E)$ with susceptances, a fixed dispatch, and a protected transfer from generation at $s$ to a critical load at $t$. The decision variables are three sets of yes/no choices: which lines to open, which lines form each of the $k$ node-disjoint corridors, and, optionally, which boundary interfaces to tune to the rank-1 condition.
Let $F_e$ be the pre-outage flow on line $e$ and $L_{ef}$ the LODF of line $e$ with respect to the outage of line $f$. The flow on $e$ after $f$ trips is then $F_e + L_{ef}F_f$. We call corridor set $\{P_1,\dots,P_k\}$ cascade-independent if, for every contingency $f$, at least one corridor $P_j$ survives two rounds. In round one, $P_j$ is not hit and no line of $P_j$ is overloaded:
$$|F_e + L_{ef} F_f| \le C_e \quad \text{for all } e \in P_j,$$
where $C_e$ is the thermal limit of line $e$. In round two, the corridor is also not overloaded after the lines that round one did overload are removed together with $f$, which uses the generalized multi-outage factors from Part I.
Topology changes alter the LODFs. We handle this with the standard rank-1 update of the Laplacian pseudo-inverse (the matrix that maps power injections to voltage angles in DC flow), applied inside a lazy-constraint loop, rather than writing the physics into the MILP in bilinear form.
The scaling rests on screening. By the simple-cycle criterion, a non-cut outage $f$ can affect corridor lines only if $f$ lies in a block that contains a corridor line. By the simple-path criterion, a cut outage matters only if it lies on a simple path from the outage boundary to buses that take part in balancing. A block-cut tree, which records how the biconnected blocks attach to each other, can be recomputed in linear time after every candidate switching, so the contingency set shrinks from $|E|$ to the lines in corridor-touching blocks. Opening lines splits blocks further. The solver therefore has a real incentive to switch, and every switch shrinks the verification problem as well.
The algorithm has three parts. A master MILP chooses openings and corridors. An oracle computes post-switching blocks and screened contingencies, runs the two-round DC cascade, and returns violated $(f,e)$ capacity constraints as cuts. We expect to use no-good cuts (constraints that forbid a rejected configuration from being proposed again) only when a switching pattern makes the LODF linearization inexact. ShareDP-style batch search warm-starts the master with candidate corridor sets.
The deliverables are an exactness result for the screening, a solver, and an AC validation study. The exactness result states that screened and unscreened formulations have the same optimum. Here it needs its own check, because the round-two cascade can change block structure mid-cascade. The AC study addresses the linearization limitation that the localization literature itself acknowledges.
Baselines
The naive reference point uses no switching. On the unmodified grid it computes the $k$ node-disjoint $s$–$t$ corridors of minimum total length, then scores them with the same two-round cascade simulator. This is what a routing engineer would deploy today, and it isolates the cost of ignoring flow redistribution.
The strongest graph-only competitor uses the same unswitched grid but picks, among ShareDP-enumerated disjoint path sets, the one with maximum path uniqueness in the sense of Chen et al.. It gets the same enumeration budget as our warm start. Beating it shows that electrical coupling, not graph overlap, is what predicts joint failure.
The strongest physics-side competitor is topology control aimed at localization. We run the tree-partitioning switching of Less is More and the bridge-block alignment of Adaptive Network Response to Line Failures on the same grids and then route the best available corridors on the result. For learned switching, we also train the physics-guided agent of Blackout Mitigation via Physics-guided RL with corridor survival as its reward, under a wall-clock budget matched to our solver.
No reported result found on Lacuna for corridor survival or cascade survival rates of any of these methods, so every baseline number will come from our own runs on shared instances, contingency lists, and simulators.
Three ablations isolate the proposal's ingredients. The first removes block-based screening and imposes all $|E|$ contingencies, which isolates the scalability contribution; optima must match. The second forbids switching, which isolates the value of topology control. The third drops the round-two constraint, which shows whether first-order LODF protection alone is enough.
To count as beating the baselines, the method must deliver a statistically clear gain in the fraction of contingencies with at least one surviving corridor over both the graph-only and the topology-control baselines, at no more than a matched increase in redispatch cost and corridor length. It must also solve to proven optimality on instances where the monolithic formulation times out.
Why not just…?
Why not just pick disjoint corridors and run a standard N-1 security check afterward? Generate-and-filter can only reject corridor sets on a fixed topology. The cycle argument says that on a meshed grid, every disjoint pair is coupled, so the filter may reject everything, or accept only long, costly detours. The no-switching ablation is exactly this alternative with an exhaustive filter. If it matches our survival rate at similar cost on realistic grids, the switching component is not worth its complexity, and we would report that.
Why not just pose it as one security-constrained transmission-switching MILP in Gurobi? This is the most serious threat on small grids, and we expect it to win or tie there. The problem is that constraints grow with contingencies times monitored lines times switching states. The screening ablation measures the crossover size directly and checks that both formulations reach the same optimum where both finish. If the monolithic model stays competitive up to utility-scale grids, the contribution shrinks to a modeling one.
Why not just partition the grid into a tree of areas, as in "Less is More"? A tree partition contains disturbances, but a protected transfer between two areas then depends on a single bridge. The file flags exactly this vulnerability to uncontrolled islanding. Tree partitioning is a baseline in our experiments. We expect it to score well on containment but poorly on corridor survival whenever $s$ and $t$ fall in different areas. The rank-1 interfaces of Interface Networks are how we keep redundancy across those boundaries without the coupling.
Proposed Experiments
The file lists no named power-grid benchmark with an artifact id. We will therefore use widely used public transmission test systems spanning roughly a hundred to a few thousand buses, with published line limits. For scaling tests on the graph side, we will add synthetic topologies from MoleNetwork. Protected transfers will be sampled between large generators and high-priority loads, with $k \in \{2,3\}$.
The primary metric is corridor survival, meaning the fraction of single-line contingencies after which at least one corridor delivers the full transfer under the two-round DC cascade. Secondary metrics are total load shed in the cascade, the number of opened lines, redispatch cost, corridor length, and time to proven optimality.
Three checks decide the outcome. First, on grids where both formulations finish, screened and full formulations must return identical optima. Any gap would falsify the screening argument once cascades change block structure. Second, the gain in corridor survival over the uniqueness-maximizing baseline must hold across sampled transfers, not just on average. Third, the chosen designs will be re-simulated under full AC power flow. The claim survives only if AC corridor survival stays close to the DC prediction; a large gap would mean the linearization, not the method, drives the result.
References
- Line Failure Localization of Power Networks Part I: Non-Cut Outages
- Less is More: Real-time Failure Localization in Power Systems
- Adaptive Network Response to Line Failures in Power Systems
- Line Failure Localization of Power Networks Part II: Cut Set Outages
- Interface Networks for Failure Localization in Power Systems
- Evaluating Distribution System Reliability with Hyperstructures Graph Convolutional Nets
- Reliable Virtual Machine Placement and Routing in Clouds
- S-MATE: Secure Coding-Based Multipath Adaptive Traffic Engineering
- ShareDP: Finding k Disjoint Paths for Multiple Vertex Pairs
- Optimal Qos-Aware Network Slicing for Service-Oriented Networks with Flexible Routing
- Blackout Mitigation via Physics-Guided RL
- MoleNetwork: A tool for the generation of synthetic optical network topologies