Forecasting at new assets: transfer learning and proxy nodes
- David Murray
- Apr 3
- 4 min read
How graph-based models and transfer learning can improve price forecasts at new nodes with limited or no historical data
The power grid is constantly evolving. In 2025, there will be 870 new assets that begin selling power to the grid, around 85% of which will be operated in a deregulated market with an associated price node. Not all may be new, but many will be. For developers and traders, forecasting prices at these new locations is critical—yet these nodes often have little or no historical data. That lack of history presents a challenge: how do you make trades for a node that’s never existed before?
Forecast accuracy at these new nodes matters. It affects how assets are valued in backtests, how traders position portfolios, and how developers negotiate off-take or optimization agreements. Any asset’s performance is a function of its price node’s exposure and the revenue capture rate of its trading operations. A poor estimate for either can skew project economics or commercial terms, even if the forecast improves over time. Machine learning techniques—particularly transfer learning within graph-based models—can help solve this challenge.
Proxy Nodes: A Common But Incomplete Fix
In early-stage development or valuation, a typical workaround is to use a proxy node—one that is geographically close to the new location. This provides a first approximation of how prices might behave at the new node. But over time, the new node's price behavior often diverges from its proxy due to localized congestion, new generation patterns (duh), or subtle differences in grid topology. Proxy nodes are a great way to assess the potential of a location on the grid, but for operational purposes can eventually feed into sub-optimal decisions.
Modelling the Grid as a Graph
In machine learning, the grid can be represented as a graph, where nodes are price locations (hubs, generators, interties) and edges reflect relationships between them. These relationships can represent physical transmission lines (and those on outage), proximity in dispatch stacks, or other inferred connections. Each node is enriched with metadata—latitude and longitude, region, asset type — much of which mirrors inputs from production cost models.
To create a graph structure that captures the most important connections without overfitting, we use a minimum spanning tree. This approach connects all mathematical nodes in the most efficient way possible, minimizing total "distance" between them. The result is a simplified yet powerful representation of how prices interact across the system. By connecting all the price nodes in a graph representation, you can use the new node’s relationship to the rest of the grid instead of to a single similar node when making predictions.
What Transfer Learning Looks Like in Practice
Transfer learning enables a model trained on one set of nodes to generate meaningful predictions for another—even if the new node lacks significant historical data. In this approach, a graph embedding model is first trained on historical price behavior across the grid. The model identifies patterns in how different types of nodes behave using factors such as geography, topology, and technology type. When a new node is introduced, it is positioned within the graph using its metadata and physical location. Its likely price behavior is then inferred based on the most similar nodes in the embedding space.
The outcome is an initial price forecast that incorporates both the structure of the grid and the historical behavior of comparable nodes.
Key Takeaways
A common approach to valuing new assets is to select a nearby price node with similar congestion and technology
Transfer learning expands on this approach by using the new nodes relationship with many other nodes on the grid
The model can be trained on the entire grid with a long history of data and use
metadata about the topology to predict the price at the new price node
Why It Matters for Developers and Traders
This approach isn’t just theoretically elegant—it has real commercial impact. By improving short-term price accuracy at new nodes, traders can make smarter early bets at their asset, developers and IPPs can tighten the gap between forecasts and realized performance and commercial teams can negotiate better off-takes, based on defensible forecasts.
Transfer learning helps eliminate the blind spot that new nodes create. When combined with careful proxy node selection, it provides a powerful framework for early-stage forecasting and valuation.
Enertel AI provides 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.
If the methodology proposed in this post could be useful to you, don’t be afraid to reach out!
Comentários