Meridia Insight Tech for Good Frontiers

A 2,000-bus Texas grid, planned one investment at a time

A 2,000-bus Texas grid model links battery siting, line upgrades and investment timing—and reports a solution within 1% of its bound.

2,000 buses, 2030–2045, and a reported gap below 1%: a model coordinates batteries and line upgrades across time.

A Texas grid model with 2,000 buses produced a coordinated transmission-and-storage investment plan for 2030–2045 with a reported optimality gap below 1%. That matters because the decisions now facing grid planners are not simply whether to build more infrastructure. They must decide where to reinforce the network, when to do it, and whether batteries can make a bottleneck less costly than upgrading a line. Those choices interact across years—and, at the scale of a real regional grid, can overwhelm the optimization tools planners use to compare them.

Kevin Wu, Rabab Haider and Pascal Van Hentenryck present a method designed to solve that coupled planning problem at high spatial and operational resolution (Wu et al., 2026). Their model weighs transmission reconductoring—the replacement of a line’s conductor to increase its carrying capacity—against installing batteries at individual grid nodes. It plans across multiple future periods rather than optimizing each period in isolation. The reported result is promising, but the paper excerpt available here does not include the detailed investment totals or numerical comparison with the rolling-horizon method. The strongest conclusion is therefore about computational feasibility: this approach can solve a large, detailed planning model to a tight reported gap, not that it establishes one universally best Texas investment plan.

The network in the case study

The paper demonstrates its method on a Texas system with 2,000 buses.

The network in the case study
LabelValue
Texas case system2,000 buses

The Science

A power grid is a physical network, but investment planning is also a sequencing problem. A transmission upgrade built early can serve the system for years; a battery may defer that upgrade, or make it more useful, by shifting electricity across hours. Meanwhile, demand and the mix of generation change. Planning each year as a separate puzzle risks missing the value—or the cost—of decisions made earlier.

The researchers formulate what they call an all-years model. It chooses the timing and location of line reconductoring and storage investments across planning periods, then tests how the resulting system would operate. The investment decisions are irreversible: a line upgrade or battery installed in one period remains available in later periods. The objective adds discounted capital and operating costs, giving later spending less weight than spending in the initial period. For operations, the model includes generation costs, battery charging and discharging, and a penalty for any electricity demand that cannot be served.

The network retains a useful degree of physical detail. The model uses a DC power-flow approximation, a common planning tool that estimates how electricity moves through a grid while simplifying some of the full alternating-current physics. It represents flows with a power-transfer distribution factor, or PTDF, matrix: a precomputed map from injections and withdrawals at grid nodes to flows on transmission lines. Reconductoring raises a line’s thermal limit—how much electricity it can safely carry—while leaving its electrical characteristics sufficiently unchanged for this fixed-network representation.

Storage is modeled as four-hour lithium-ion batteries. Their energy capacity determines how much they can store, and their charge and discharge power are capped at one-quarter of that capacity. The model tracks battery state of charge hour by hour, including efficiency losses. But each representative operating day begins and ends with an empty battery. That makes days independent: storage can shift power within a day, but it cannot carry a surplus from one day into the next.

To represent the variety of operating conditions without simulating every day of every future year, the model uses representative days with assigned probabilities. It includes hourly operating decisions for each such day and planning period. This is substantially richer than a single snapshot of system demand, but the quality of the answer still depends on how well those representative days capture difficult conditions such as prolonged low renewable output or unusual demand peaks.

The computational challenge is that investment choices are discrete—reconductoring a particular line is a yes-or-no decision—while hourly operating choices multiply across nodes, periods and days. The authors tackle this with Benders decomposition. In plain language, it splits the big problem into a master task that proposes investments and smaller operating tasks that test whether those investments work and what they cost. The operating tasks send mathematical feedback, called cuts, back to the investment task. Because each period and representative day can be evaluated separately, many of those checks can be run in parallel.

This kind of iterative method can wobble. One investment proposal may look attractive, the operating model may expose a costly bottleneck, and the next proposal may swing too far in the other direction. The authors stabilize the search with a trust region: each iteration is encouraged to stay within a bounded distance of the best investment plan found so far. Crucially, they impose separate bounds for each planning period, rather than treating the entire multi-decade trajectory as one undifferentiated decision.

They also use a warm start. Before the main iterative solve, they relax the investment problem, temporarily allowing fractional decisions, and use those results to construct a plausible starting plan. The model takes the most demanding relaxed investment suggested by the representative days and rounds to an allowed investment level. It also narrows the candidate lines and nodes to those flagged by the relaxed solutions. These choices help the solver start near useful options rather than at a zero-investment plan that leaves demand unserved.

What They Found

The headline demonstration is a 2,000-bus Texas system, with an investment horizon from 2030 to 2045. The authors report that their plan reaches an optimality gap below 1.0%. An optimality gap measures how far the best feasible solution found is from the strongest lower bound the solver can establish; a small gap gives evidence that the computed plan is close to the model’s best possible answer. It does not mean the plan is guaranteed to be best under every forecast or modeling assumption.

That distinction matters. The result is not a forecast that Texas will build a particular number of batteries or upgrade a particular set of lines. It is a computational result about what the model can solve given its inputs, constraints and representation of future conditions. The supplied paper text does not report the final investment quantities, total system costs, numerical runtime, or the numerical outcome of the rolling-horizon comparison. Those figures should not be inferred from the abstract’s headline.

The paper’s first figure tracks the optimality gap as runtime accumulates, including the warm-start phase. It is the right place to see how quickly the algorithm closes the distance between a feasible investment plan and its bound, though the graph’s underlying point values are not included in the text supplied for this digest.

Fig. 1: AY TR-Benders optimality gap over total runtime (including WS)
Fig. 1: AY TR-Benders optimality gap over total runtime (including WS) Source: Kevin Wu, Rabab Haider

The second figure separates the time spent in the master problem from the time spent in operating subproblems. That breakdown matters for future scaling: if operations dominate, more parallelism or faster operational checks may help; if investment decisions dominate, the master problem may need further strengthening. Again, the available text provides the figure but not a table of its plotted values.

Fig. 2: AY TR-Benders MP and SP computation times
Fig. 2: AY TR-Benders MP and SP computation times Source: Kevin Wu, Rabab Haider

The model explicitly compares its all-years planning approach with a rolling-horizon approach. Rolling-horizon planning repeatedly solves a near-term problem, makes the next investment decision, then updates and solves again. It is intuitive and can be practical, but decisions made today may not account fully for their consequences across the whole future horizon. The authors use the comparison to highlight the potential value of coordinating investment timing across periods. Without the comparison’s numerical results in the supplied material, however, it is not possible to say how much the coordinated plan saves or which specific investments differ.

Battery duration represented

The model represents lithium-ion storage with four-hour duration.

Battery duration represented
LabelValue
Modeled battery4 hours

Why This Changes Things

The central advance is not that batteries or transmission upgrades are new. It is that the authors bring three planning dimensions together at a scale where their interaction matters: location across a detailed network, timing across multiple investment periods, and hourly operation under different representative conditions. A battery placed on the wrong side of a congested line may not relieve the constraint planners care about. A line upgrade may be less valuable if storage can absorb a local renewable surplus and discharge during a later peak. Looking at either asset class alone can miss those substitutions and complementarities.

That is increasingly relevant as electricity demand grows from electrification, manufacturing and data-intensive computing, while renewable generation adds variability in where and when power is available. Grid infrastructure is slow to plan and build. Better ways to compare options could help system planners ask not just “what is needed?” but “what must happen first, and what can wait?” This paper offers an optimization framework for asking that question with explicit network flows rather than a simplified regional or single-node representation.

A near-1% reported gap is also meaningful as a tool for decision-making. Large planning models often return a plausible plan without showing how close it is to the best solution the mathematical formulation permits. A tight bound gives planners and analysts more confidence that the algorithm did not stop at an obviously inferior corner of the search space. It is not a guarantee that the assumptions are correct, but it helps separate solver uncertainty from uncertainty about the future.

The method’s design choices may also travel beyond Texas. Per-period trust regions use the structure of a multi-year investment problem instead of applying a generic stabilization rule. Parallel subproblems offer a natural route to using modern computing resources. And the relaxation-based warm start could be useful whenever a large investment model has many discrete choices and is difficult to launch from a feasible plan.

Reported optimality-gap threshold

The abstract reports an optimality gap below 1.0%; the bar shows the stated 1.0% threshold, not an exact measured gap.

Reported optimality-gap threshold
LabelValue
Reported threshold1 % maximum reported

Still, more modeling detail does not remove uncertainty. The DC approximation does not capture every feature of real alternating-current power flows or all reliability concerns. Reconductoring is the only transmission intervention represented here; entirely new lines and other grid technologies are outside the formulation. The batteries are four-hour lithium-ion devices, so the model does not compare long-duration storage technologies. Its day-by-day storage boundary also rules out moving energy across days, a potentially important capability during multi-day weather events.

And a close solution to an optimization model is only as decision-relevant as the scenarios and costs fed into it. Demand growth, fuel prices, renewable availability, project costs and permitting delays can all shift the answer. A robust planning tool should ultimately test how investment choices change across plausible futures, not only optimize one modeled set of assumptions.

What's Next

The next test is whether the approach remains useful as planners add the complications that shape real investment decisions: uncertainty in future demand and generation, contingencies that test reliability, new transmission corridors, more storage durations, and energy shifting across multiple days. The authors’ formulation is a foundation for such work, not a complete digital replica of the grid.

The rolling-horizon comparison also deserves a fuller accounting. How much does coordination change costs or investment timing? Which upgrades are brought forward, and where does storage substitute for transmission—or the reverse? Answering those questions with the full numerical results would help planners judge whether the extra computational effort produces material system benefits, not only a tighter optimization certificate.

For now, the notable step is one of scale and coordination: a detailed Texas network, multiple planning periods, hourly operations, transmission upgrades and storage investments all in a single optimization framework, with a reported gap below 1%. If that combination can be made transparent, stress-tested and fast enough for repeated scenario analysis, it could help grid planners evaluate infrastructure choices as a connected sequence rather than a series of disconnected bets. That is where better computation can begin to matter beyond the solver: in helping societies build a grid that is more reliable, flexible and ready for an uncertain energy future.

Comments (0)

No comments yet. Be the first to share your thoughts.