Rabab Haider, Assistant Professor of Civil and Environmental Engineering, College of Engineering
Watch RecordingRabab Haider opened her talk with a problem most people encounter only as a high electricity bill or a brief outage: congestion on the power grid. The grid was designed for a world of stable, predictable loads and centralized generation. That world no longer exists. Data center demand within MISO’s footprint alone is projected to grow 130% by 2027. Wind generation in Texas is arriving at the wrong end of the grid, far from the load centers it needs to serve. The infrastructure cannot be rebuilt fast enough — “rebuilding the airplane while it’s in flight,” as Haider put it — so the question her group has been asking is whether the existing grid can be operated more flexibly.
The answer lies in network reconfiguration: opening and closing transmission switches, and performing what engineers call bus splitting — electrically decoupling a single substation into two independent nodes, effectively adding a node to the network graph without any physical construction. These are software actions, not steel ones, and they can meaningfully expand the range of loads a grid can support. Haider’s simulations on a standard IEEE test system showed that reconfiguration alone extended the feasible operating region, and that bus splitting extended it further — allowing data center loads that the base topology simply could not support.
The hard problem begins at scale. The standard test systems have 30 buses. The open-source synthetic model of the US grid has 82,000. Scaling from 30 to 82,000 requires replacing the slow mixed-integer optimization solver — which finds the optimal topology but takes far too long for real-time grid operations — with something faster. Haider’s group turned to neural networks as optimization proxies: trained to approximate the solver’s output, they can evaluate candidate topologies in milliseconds rather than hours. But neural networks operating on physical systems introduce a class of problem that has no analogue in standard machine learning: the predictions can be physically impossible. A network that has been trained to predict switch configurations might output a topology that violates the laws of electrical physics. Worse, the switching variables are binary — on or off — and neural networks have no native mechanism for enforcing integer constraints.
The majority of Haider’s AI journey has been spent navigating that gap between what neural networks can learn and what the physics demands. Three families of approaches have emerged from that work. The first exploits operational knowledge: incorporating constraints from industry partners like Siemens and RTE — the French system operator — that reflect real-world protection schemes, market design rules, and regulatory requirements that would never appear in a physics textbook but govern how actual operators behave. The second uses repair layers: taking a physically infeasible neural network output and projecting it toward the nearest feasible solution using power-flow algorithms. The third encodes the network’s physical structure directly into the learning architecture using graph neural networks with physics-aware edges, giving the model structural knowledge of the system it is operating on rather than treating the grid as an undifferentiated set of inputs.
The clearest validation came from a problem that Gurobi — the best commercial mixed-integer solver available — could not solve within operational time limits. Haider’s heuristic optimization proxy produced a feasible solution where the conventional solver timed out. That result — a problem solved, not just a model that scored well on a test set — is the kind of output that matters when the audience is a grid operator, not a machine learning researcher. Conversations with MISO are ongoing, and Haider was candid about the challenges that remain: protection schemes tied to fixed topologies, pricing structures that assume stable nodal identities, and the regulatory landscape surrounding any change to grid operations. The engineering gap between a working research system and a deployable operational tool is real, and her group is working on it alongside academic and industry partners.