When the power grid guesses wrong, the consequences are not academic. A forecast that runs a few percent high leaves utilities paying for generation nobody needs. A forecast that runs low — especially during a heat wave, a cold snap, or an unexpected surge from a hyperscale data center — can mean brownouts, emergency spot-market purchases, or worse. Most modern load-forecasting models are confident either way. They hand operators a single number and stay silent about how much to trust it.
A new paper, "EnergyMamba: An Uncertainty-Aware Graph-Enhanced Selective State Space Model for Energy Consumption Prediction," accepted at ICML 2026, takes direct aim at that silence. It pairs a fast, modern sequence model with a method for flagging its own uncertainty — and it does so in a way tuned to how electricity actually moves through a grid.
What a "selective state-space" model is
For the past several years, the default architecture for sequence problems — text, audio, time series — has been the Transformer, whose attention mechanism compares every point in a sequence against every other point. That comparison is powerful but expensive: cost grows quadratically with sequence length, which makes very long histories awkward to model.
Mamba-style selective state-space models (SSMs) offer a different bargain. Instead of comparing all pairs of timesteps, an SSM carries a compact internal "state" forward through the sequence, updating it step by step — closer in spirit to a classical recurrent network, but with a mathematical structure that scales linearly with sequence length and runs efficiently on modern hardware. The "selective" part is the key innovation: the model learns, at each step, what to remember and what to discard, letting it focus on the signals that matter and forget the noise. For energy data — long, regular streams of consumption readings shaped by daily and seasonal rhythms — that linear-time efficiency over long horizons is a natural fit.
Why grids need graphs, and why they need doubt
EnergyMamba's authors argue that treating load forecasting as a pure time-series problem misses half the picture. Electricity is a networked phenomenon. The framework's first component, a Graph-Enhanced Selective State Space Model (GE-Mamba), uses a graph convolutional network to extract spatial context from the grid's topology and injects it directly into the SSM's temporal modeling.
The motivation the paper cites is physical, not merely statistical: Kirchhoff's circuit laws. A change in consumption at one node propagates to connected nodes through the power-flow equations, so conditioning each region's temporal dynamics on its neighbors' state should yield more accurate predictions than modeling each meter in isolation. The graph encodes which regions are electrically adjacent; the SSM tracks how their demand evolves over time. The two are coupled rather than bolted together.
The second component is where the headline promise lives. The Adaptive Sequential Conformalized Quantile Regression (AS-CQR) module produces calibrated prediction intervals — ranges that are statistically guaranteed, under conformal-prediction theory, to contain the true value at a chosen confidence level. Crucially, the paper distinguishes between two kinds of uncertainty an energy system faces. Aleatoric uncertainty scales with load magnitude, because physical losses grow with throughput; AS-CQR handles it with scale-invariant normalization. Epistemic uncertainty arises from distribution shift — the unusual conditions of an extreme-weather event the model has rarely seen — and AS-CQR adapts to it through an online feedback loop that adjusts intervals as new data arrives. The result, in plain terms, is a model that widens its error bars precisely when it should be least sure.
The numbers, and the caveats
The authors report evaluating EnergyMamba against 15 state-of-the-art baselines on four large-scale real-world datasets drawn from Florida, New York, and California — utility regions chosen to span very different climates and demand profiles. Across that benchmark, they report roughly a 5 percent improvement in point-prediction accuracy and about a 6 percent improvement in uncertainty quantification over the strongest prior methods. Those are meaningful gains in a field where single-digit percentages translate into real money and real reliability margins, though as with any single paper they await independent replication.
Why this matters now
The timing is hard to ignore. Grid operators across the United States are confronting a load picture that has been roughly flat for two decades suddenly bending upward, driven in large part by AI data centers whose power draw can rival a small city's and ramp with little warning. That demand is exactly the kind of non-stationary, hard-to-model signal that breaks conventional forecasters — and exactly the regime where a model that admits its own uncertainty is most valuable. An operator who knows a forecast is shaky can hold more reserve generation, pre-position resources, or hedge in the market. An operator handed a falsely confident number cannot.
EnergyMamba's combination of linear-time efficiency, grid-aware structure, and calibrated doubt reads as a deliberate response to that environment. Selective state-space models bring the speed to handle long, high-resolution streams from thousands of nodes. The graph layer encodes the physics of how those nodes interact. The conformal layer turns a forecast from an assertion into a probability statement an operator can plan around.
What to watch next
The open questions are practical ones. Conformal guarantees are only as good as the assumption that future data resembles past data — the very assumption that breaks during the extreme events where reliable intervals matter most, and the reason AS-CQR's online adaptation will draw scrutiny. It also remains to be seen how the approach holds up on grids with rapidly added data-center load, where the historical record offers little precedent. And there is the question of deployment: research benchmarks are not control rooms, and utilities adopt new tooling slowly and conservatively. Still, the direction is clear. The next generation of load forecasters will likely be judged not only on how accurate they are, but on how honestly they say when they don't know.
"A change in consumption at one node propagates to connected nodes through the power-flow equations, so conditioning each region's temporal dynamics on its neighbors' state should yield more accurate predictions than modeling each meter in isolation."— EnergyMamba paper, method rationale (paraphrase)