Power Grids are Graphs

How Graph Neural Networks Unlock the Future of Forecasting on The Electric Power Grid

New York ISO (NYISO) represented as a mathematical graph with price nodes and connections between them.

Lately, it seems like everyone is talking about electricity prices.

Whether that's data centers driving them up, renewables driving them down, batteries flattening them out, or any combination thereof, power prices have become a central point of discussion in industrial policy circles. Constrained supply chains and growing interconnection queues have some turning their attention toward space-based compute. We're just a few short years into latest AI boom, and we're already considering launching entire solar farms on rockets to avoid connecting to the grid. Think about that for a moment.

In many ways, traditional forecasting demands the opposite: stable patterns, historical trends, clear signals. We want our models to capture behaviours of the past so that we can accurately project them into the future, but in a world where tomorrow looks so different from today, has forecasting the future of the power grid become an impossible challenge? We've spent the last four years thinking about exactly this problem, and we're happy to report some good news.

Discovering Price Networks

If you've ever looked at electricity prices across different locations on the grid, you've probably noticed something: prices at nearby nodes tend to move together. That's because electricity prices don't exist in isolation, they're set by a market-clearing process that simultaneously considers supply, demand, and transmission constraints across the entire network. When a transmission line hits its limit, price differences emerge on either side. When a generator trips off, its impact ripples through the network.

Power markets have strict timelines, so this price setting process has to approximate the real physics of the grid to avoid solving an otherwise complex, non-convex optimization problem (DCOPF vs ACOPF). These simpler equations have analogies to well-understood DC circuit laws, providing an intuitive but slightly incorrect understanding of power flow. Less obvious though, is that the prices themselves obey the same circuit laws; modelling electricity prices as voltages and their differences as flows allows for a physical interpretation of price behaviour. This finding suggests a simple but powerful idea: the statistical similarity between electricity prices encodes important information about the underlying network topology.

By replacing price differences with statistical likeness metrics and applying algorithms from graph theory, we can extract a sparse graph structure that captures some essence of the physical power system. This idea can be further extended to the temporal domain by repeating the operation over shorter date ranges, effectively growing the graph over time by creating a series of snapshots that describe an evolving and dynamic power system. By incorporating the natural growth of both supply and demand in the past, we're able to more accurately account for systemic changes in the future.

The Houston Load Zone with connections to local generator price nodes.

Adding the Physical Domain

How far can we take this idea? Price nodes are a financial construct in power markets, but they are "virtually" connected to physical power plants and electrical buses, which are connected to substations, which themselves are connected by transmission lines. A full list of planned, existing, and retired power plants by capacity, fuel type, and technology is available through the Energy Information Agency. Asset-level generation for thermal power plant data is available through open environmental reporting standards, and actual demand and renewable generation are broadly reported in aggregate values by load zone or geographic weather region. If we want to enrich our graphs with additional physical information, how should we go about doing that?

One option is to include these as new attributes on existing nodes and edges in the graph. In that case, we're left with essentially the same graph where all the nodes and edges are of the same type - price nodes with simple connections between them. The other option is to add new node and edge types entirely. We can add power plant nodes and connect them to their respective price nodes. Or we can add substation nodes and connect those to their respective load zones. In either case, we're left with a dynamic heterogenous graph that combines both physical and financial information and captures complex relationships between them.

A graph neural network of USA's largest market, Midcontinent ISO, where congestion is accurately captured by incorporating the natural growth of both supply and demand across the network.

At Enertel, we call them Power Graphs, a nod to the benchmark that inspired them and the the newest addition to our growing family of models for the electric power sector. Built on state-of-the-art graph neural networks and capable of forecasting diverse node and edge types, these models learn to exploit complex graph structures naturally contained in power systems data to deliver high performance probabilistic forecasting at the nodal level.

Existing users will gain access to the ability to create custom inference scenarios for comparison against base conditions, and in the coming weeks will be able to alter the graphs themselves in a new on-demand API inference service. Subsequent releases will introduce our first combined Dispatch and Price Signal (DPS) models, and research is underway to introduce physics-informed losses directly into the training process to facilitate long-range price forecasting on unseen grid topologies.

We’re particularly excited about the prospects of self-supervised learning for the full AC optimal power flow problem and expect to have more to say on this in the year ahead.

The Future of Forecasting

All of this points to a simple truth: this is an unusually exciting moment for the electric power industry. The rapid rise of AI is colliding directly with wholesale power markets, forcing the industry to revisit long-standing assumptions about traditional forecasting methods and optimization techniques. The graphs are getting richer and the questions more ambitious.

The future of forecasting on the electric power grid is being rewritten right now and the outlook is bright.

Enertel AI uses graph neural networks to provide short-term energy and ancillary price forecasts for utilities, independent power producers (IPPs), and asset developers to inform trading strategies. You can view our data catalog, view a clickthrough demo of our product for operators, or request a sample of backcast data for your assets.