Nodal Congestion Forecasting: Predicting When and Where Transmission Constraints Bind

Why prices split across locations

In organized wholesale electricity markets, every pricing node on the grid receives a locational marginal price (LMP): the marginal cost of delivering one additional megawatt of energy to that specific location. When the transmission network has spare capacity everywhere, power flows freely and all nodes see roughly the same price. But transmission lines and interfaces have physical limits. When a constraint binds, meaning a line or interface reaches its limit and the market operator must redispatch generation to keep flows feasible, prices diverge. Nodes behind the constraint, where cheap power cannot reach, see prices rise. Nodes on the export side see prices fall. This price separation is the congestion component of the LMP.

Congestion is not a local phenomenon, and that is what makes it hard to forecast. A nodal price is not simply a local time series. Congestion emerges from the interaction of the entire network state: the pattern of injections (generation) and withdrawals (load), transmission outages, and contingency limits (limits enforced so the system survives the loss of an element, such as a line or generator). The mapping from network state to prices is discontinuous. A small change in load, generation, or topology can switch which constraint is marginal and reprice a large region of the grid at once. Standard time-series methods, which assume smooth local dynamics, struggle with exactly this behavior.

Congestion is also where much of the forecasting value sits. Financial transmission rights (FTRs) and congestion revenue rights (CRRs), instruments that pay out based on congestion price differences between a source and a sink node, settle entirely on this component. A model can achieve low average nodal-price error while failing on the few constraints that drive most of the congestion value.

What to predict

Congestion forecasting is not a single target. Useful prediction targets include:

• the probability that a monitored constraint binds;
• the shadow price of each constraint (the marginal value of relaxing its limit by one unit);
• the congestion component of the nodal LMP;
• the source to sink congestion spread between two nodes;
• the active-set ranking, meaning which constraints are likely to be binding simultaneously;
• the topology or outage regime the system is operating in;
• the payout distribution of an FTR or CRR position.

Framing matters. Predicting which constraints bind, then deriving price implications, is often more tractable than predicting nodal prices directly.

Classical approaches

The classical toolkit is grounded in power-system physics:

• DC power-flow simulation with power transfer distribution factors (PTDFs) and line outage distribution factors (LODFs). PTDFs are linear sensitivities describing how an injection at one node changes flow on each line; LODFs describe how flow redistributes when a line goes out.
• Contingency analysis, which checks feasibility under hypothetical element losses.
• Linear sensitivity models built on those shift factors.
• Logistic regression or boosted-tree classifiers that predict whether a given constraint binds, using features such as load, generation, and outage patterns.
• Nearest-neighbor matching of historical load and outage patterns to find analogous days.
• Scenario-based production-cost modeling, which simulates market dispatch over many scenarios.

These methods remain strong precisely because they preserve physical meaning. A PTDF-based simulation with boosted-tree binding classifiers on top is a credible, well-understood baseline, and in practice it is the recommended starting point.

Graph learning

A power network is naturally a graph: buses (nodes) connected by lines and transformers (edges). Graph neural networks (GNNs), models that learn by passing messages along graph edges, are a natural fit. Approaches in use include graph convolutional networks, graph attention networks, message-passing neural networks, spatiotemporal GNNs that model both network structure and time, line-graph models that place branch constraints on nodes of a transformed graph, and heterogeneous graphs with distinct node types for buses, generators, loads, and interfaces.

Node features can include load, generation, weather, and historical prices. Edge features can include line reactance, thermal limits, outage state, and flow. A GNN may predict LMPs directly, or it may predict the set of active constraints and let a physics layer compute the price implications. The second design is generally safer, for reasons covered next.

Physics-informed and hybrid models

The highest-value designs combine learning with optimization rather than replacing it:

1. Active-set prediction: a machine learning model ranks the constraints most likely to bind, and an optimal power flow (OPF, the optimization the market operator solves to dispatch generation at least cost subject to network constraints) verifies them.
2. Residual correction: a DC-OPF produces a base congestion forecast and the ML model estimates the residual error.
3. Differentiable optimization: the model is trained end to end through an OPF layer.
4. Topology embeddings: switching and outage configurations are represented explicitly as inputs.
5. Constraint-conditioned price models: the model first classifies the constraint regime, then predicts shadow prices conditional on it.

A consistent design principle applies across all of these: the optimizer should remain the feasibility authority. A GNN can accelerate constraint screening or improve an approximation, but direct nodal-price predictions from a neural network can violate network logic, producing price patterns no feasible dispatch could generate.

The data problem

The main bottleneck in congestion forecasting is not model capacity. It is reconstructing the market network as it was known at bid time. That reconstruction requires:

• planned and forced transmission outages;
• temporary ratings and ambient-adjusted limits (line limits that change with weather conditions);
• nomograms and interfaces (operator-defined composite limits over groups of lines);
• contingency definitions;
• phase shifter and HVDC control settings;
• transmission switching actions;
• operator-entered constraints;
• the mapping between public node names and the underlying network buses.

Much of this information is restricted or incomplete. Detailed network models are typically covered by Critical Energy Infrastructure Information (CEII) rules, which limit public access. The binding-constraint files that market operators publish describe outcomes, not necessarily the complete set of inputs that produced them. And the topology itself changes over time, so a network model assembled today may not describe the grid as it existed in the training data.

Evaluation and failure modes

Evaluation should measure both classification quality and economic performance: precision, recall, and PR-AUC for rare binding events; top-k recall on the active constraint set; error on shadow prices and congestion components; sign and rank accuracy for source to sink spreads; FTR/CRR portfolio profit and drawdown; robustness to outages not seen in training; and physical consistency under power-flow checks.

Recurring failure modes include training on a fixed graph while the real topology changes, conflating statistical correlation with true shift-factor exposure, using finalized outage data that was not available at auction time, ignoring contingency constraints, predicting each node independently rather than jointly, evaluating on a random node split that leaks structure from neighboring nodes, and assuming public constraint names are stable identifiers across time.

Practical maturity

The approaches sit at different maturity levels. Classical shift-factor simulation, contingency analysis, and boosted-tree binding classifiers are mature and remain strong baselines because they encode the physics directly. Graph learning is promising but is best deployed in hybrid form. The current frontier is a topology-aware spatiotemporal GNN that predicts active-set and shadow-price distributions, coupled to OPF validation for feasibility and an uncertainty-aware source to sink portfolio optimizer for decision making. In every configuration, the physics-based optimizer, not the learned model, should have the final word on feasibility.

Glossary

• LMP (locational marginal price): the marginal cost of delivering one more megawatt to a specific grid location.
• Binding constraint: a transmission limit that is reached, forcing redispatch and causing prices to separate across locations.
• Shadow price: the marginal value of relaxing a binding constraint's limit by one unit.
• PTDF (power transfer distribution factor): the sensitivity of flow on a line to an injection at a node.
• LODF (line outage distribution factor): the sensitivity describing how flow redistributes when a line goes out of service.
• OPF (optimal power flow): the optimization that dispatches generation at least cost subject to network constraints.
• Contingency: a hypothetical loss of a system element that the operator must remain secure against.
• FTR/CRR (financial transmission right / congestion revenue right): financial instruments that settle on congestion price differences between a source and a sink node.
• Active set: the set of constraints binding simultaneously in a given dispatch.
• CEII (Critical Energy Infrastructure Information): regulatory restrictions limiting public access to detailed network model data.
• GNN (graph neural network): a neural network that learns by passing messages along the edges of a graph.