Electricity users are increasingly asked to cut "their" emissions, and the most defensible way to define "their" is consumption-based accounting. This method traces each tonne of CO2 from the power plant that emitted it, through the transmission network, to the load that drew the power. Recent graph algorithms make this tracing fast enough to run every few minutes on an 8,870-bus model of California. This note asks what happens once that ledger becomes a target instead of a report. A region or load aggregator that schedules flexible demand to lower its consumption-based footprint can do it two ways. It can make the system burn less fossil fuel. Or it can draw the clean electrons that some other region was already receiving and leave the dirtier ones to its neighbours. Both lower the region's ledger, but only the first lowers emissions. The same split decides whether a region's apparent "decoupling" of economic output from emissions is real.
The proposal does two things. First, it measures how much of the reduction achieved by carbon-aware load control is real and how much is reshuffling. Second, it builds an online controller that meets a consumption-based carbon budget while explicitly limiting the emissions it pushes onto other regions, and that is cheap enough per decision to run across thousands of buses. The controller extends recent constrained online learning, which handles unknown horizons and adversarial constraints, by replacing its expensive projection step with a cheaper linear one. That step is only a contribution if it pays off, and the experiments are designed to show whether it does.
Core Claim
Consumption-based carbon accounting was built to fix a measurement problem: territorial inventories let importing regions look clean (Kang et al. estimate that 35% of North China's consumption-based carbon responsibility in 2020 would come from imported power). Our claim is that this accounting, when used as an online control objective, creates a second illusion of its own.
The mechanism is a conservation property. With line losses assigned to loads, as Yang et al. argue is necessary, the consumption-based totals of all regions sum exactly to system emissions. Any drop in one region's total that does not appear as a drop in system emissions must therefore show up as an increase somewhere else. We define the displaced share of region $r$'s reduction as
$$\lambda_r = \frac{\sum_{r' \neq r} \Delta \mathrm{ACN}_{r'}}{-\Delta \mathrm{ACN}_r},$$
where $\mathrm{ACN}_r$ is the amount of carbon consumption attributed to region $r$ over the evaluation window and $\Delta$ denotes the change relative to a cost-only dispatch. The value $\lambda_r = 0$ means the whole reduction is real. The value $\lambda_r = 1$ means it is pure reshuffling, and $\lambda_r > 1$ means system emissions rose while the region's ledger fell.
We predict that controllers steering on locational average emission rates will show large $\lambda_r$ in transmission-constrained hours. These are the signals most carbon-aware scheduling uses, and they are what Shen et al. compute in linear time. The reason is that a clean average rate at a node says nothing about which generator responds to an extra megawatt there. Shen et al. found that marginal rates can even be negative while average rates vary up to 100x across counties, so the gap between the two is not a corner case.
The constructive claim is that adding a running cap on $\sum_{r' \neq r}\Delta \mathrm{ACN}_{r'}$ to the region's budget fixes this cheaply. The cap enters as a second constraint in a constrained online learner. The region still pays for its consumption-based footprint, which is the fairness property that motivates the accounting. But it can no longer meet its budget by exporting dirt. This gives the "dual-control" idea from the decoupling literature a concrete form: Ke et al. argue that strong decoupling needs simultaneous caps on total and intensity. Here both caps are expressed in consumption terms, and a leakage guard is added.
Research Program
Accounting layer. We reuse the Tarjan/Kahn tracing of Shen et al. as given. It works by collapsing loops in the power-flow graph into single nodes and then sorting the result in topological order, which yields per-node average and marginal emission rates in time linear in network size. We add a regional aggregation layer and loss allocation to loads. These two additions make conservation hold exactly, so $\lambda_r$ is well defined.
The environment is a merit-order DC optimal power flow re-solved each interval. Optimal power flow is the dispatch problem that picks generator outputs to meet demand at least cost within line limits; the DC version linearizes the physics. This choice inherits the DC simplification that Nan et al. also make. We accept it because Shen et al. report that the full AC solve on the California system takes hours. We also check on the 118-bus system that the sign and rough size of $\lambda_r$ survive an AC re-solve.
Controller. The decision at each interval is the vector of flexible load and storage setpoints across the region's buses over a short look-ahead window. The feasible set is a polytope: bus-level bounds, storage state-of-charge chains, and a regional energy-service requirement. The learner sees the per-interval cost and two constraint signals after acting. The first is the region's consumption-based emissions against a budget path. The second is the increase in other regions' consumption-based emissions. Renewable output is drawn from realistic scenario sets and treated as adversarial.
The starting point is the anytime constrained online learner of the ICML 2026 COCO paper. It tracks cumulative constraint violation with a time-varying exponential penalty, which gives $O(\sqrt{t})$ regret and $\tilde O(\sqrt{t})$ cumulative violation without knowing the horizon. Its stated limitation is that every step still projects onto the base set. We replace that projection with a linear-minimization step in the style of Frank-Wolfe, which moves toward a vertex of the feasible set instead of solving for the nearest feasible point. This follows the online conditional-gradient methods of Wan and Zhang and Wan, Xue, and Zhang.
On our polytope, the linear step splits across buses into small path-structured linear programs plus one coupling multiplier, so it is much cheaper than a coupled quadratic program. The cost is that the set has flat faces. Wan and Zhang's optimal projection-free rates require curved, strongly convex sets. We therefore do not expect to match $O(\sqrt{t})$. The theoretical goal is an anytime bound with sublinear regret and sublinear cumulative violation of both constraints, at a weaker rate characterized by how many linear steps are taken per round. Running several conditional-gradient steps per round, as OCG+ does, is the knob that trades compute for rate.
Decoupling read-out. For each region and simulated year we compute the decoupling elasticity of Meng et al.: percentage change in emissions divided by percentage change in economic output. Its sign and size place the region in one of the Tapio states, such as strong decoupling (output up, emissions down) or weak decoupling (both up, output faster). We compute it three ways: with territorial emissions, with consumption-based emissions, and with system-consistent emissions that add the region's displaced share back in. Output is approximated by served energy weighted by load-class value, which is a crude proxy that we state openly.
The analytic point is that the Tapio state is a threshold on a ratio. When output growth is small, small errors in emissions growth flip the verdict. We therefore report a margin: how large $\lambda_r$ must be to change a region's state from strong to weak decoupling or worse. This connects to the warning from Huang, Zhou and Cheng that scale growth routinely overwhelms intensity gains, and to the rebound instability Meng et al. document. Here the "rebound" is spatial rather than behavioural.
Baselines
The naive reference is cost-only dispatch with no carbon objective. Every $\Delta$ above is measured against it. A second reference is ex-post tracing only, the use case of Shen et al.: emissions are measured and billed but never acted on. The question for this reference is whether billing alone changes anything when loads cannot react within the interval.
The most common practical competitor is greedy average-rate shifting. Each interval, flexible load moves to the buses and hours with the lowest locational average emission rate, subject to the same polytope. It is the method our core claim predicts will leak. Beating it means a comparable cut in the region's consumption-based emissions with $\lambda_r$ at least halved, at equal or lower cost.
On the 39- and 118-bus systems we also run the source-load bilateral carbon-trading mechanism of Nan et al., with its fixed 50/50 responsibility split and Aumann-Shapley allowance allocation. These are the strongest reported carbon-reduction numbers on these test systems in the file: a 16.0% reduction on the modified 39-bus system and 16.8% on the modified 118-bus system, against 4.6%/6.3% for load-side-only trading and 9.7%/9.2% for source-side-only trading. We re-implement it on the same networks, storage fleet and renewable scenarios. We then report both its system reduction and its $\lambda_r$, which that paper never measures. Beating it means matching its system-level reduction in an online setting with no forecast of renewables.
The strongest optimization baseline is the projected anytime learner of the ICML 2026 COCO paper, given the same two constraints and the same feedback. Its paper reports theoretical $O(\sqrt{t})$ regret and $\tilde O(\sqrt{t})$ violation, plus a qualitative improvement over the doubling trick on the RIPE Atlas routing benchmark. No scalar result for it was found on Lacuna. We give it the same per-interval wall-clock budget and the same tuning grid for step sizes and penalty schedule. Beating it means equal regret and violation at lower per-interval compute, or better regret and violation at equal compute.
An offline hindsight linear program that knows the year's renewables sets the upper bound on what any online method could achieve.
Two ablations isolate the proposal's ingredients. The first removes the leakage constraint and keeps only the consumption-based budget; this tests whether the guard, not the online learner, is what drives $\lambda_r$ down. The second swaps the linear step back to a projection; this tests whether the projection-free step buys anything at the tested scales.
Why not just…?
Why not just steer on marginal emission rates and drop the consumption-based budget? This is the most serious threat. Marginal rates measure the system effect of each megawatt directly, so a marginal-rate controller should have $\lambda_r$ near zero by construction. Our reply is partly conceptual. Marginal rates do not sum to total emissions, can be negative (Shen et al.), and so cannot serve as a responsibility ledger or a budget a regulator would sign. The reply is also partly empirical, and we run a marginal-rate controller as a full arm. If it meets the regional consumption budget and achieves the same system reduction as our guarded controller, the leakage constraint is redundant, and we will report the result as "use marginal signals for control, average signals for billing."
Why not just solve a forecast-driven optimization each interval with a commercial solver? At DC scale a linear program over California's network solves in seconds, so model-predictive control with renewable forecasts is feasible. Its guarantees depend on forecast quality, and it gives no violation bound when forecasts are wrong. We run it with forecasts sampled from the renewable scenario sets and compare cumulative budget violation and $\lambda_r$ under forecast error. If it dominates even at realistic error, the online-learning component is not needed and the contribution reduces to the leakage measurement and guard.
Why not just keep the projection in the ICML 2026 learner? Projection onto our polytope is a coupled quadratic program, not an SVD. The cubic-cost argument from matrix completion (Wan and Zhang) does not carry over automatically, and flat faces cost projection-free methods convergence rate. This is a genuine possibility, and the projection ablation settles it on the 118-bus system, a mid-size California region, and the full California system. If projected updates fit the interval budget at all three scales with better violation, we say the projection was not the bottleneck here and drop that claim.
Why not just use annual input-output decoupling models? Models like Ke et al. work on annual provincial tables. They cannot see interval-level redistribution among regions sharing a grid, which is exactly where $\lambda_r$ lives. They remain the right tool for cross-sector questions, and we do not compete with them there.
Proposed Experiments
The main testbed is the California Test System from Shen et al.: 8,870 buses, 2,149 generators and 2019 CAISO-based dispatch, whose traced totals they report within 6% of official CAISO values. We partition it into county-level regions and let one or several coastal regions act as controllers. Flexible-load shares and storage capacities are swept.
Mid-scale tests and the comparison with Nan et al. use the modified IEEE 39- and 118-bus systems. Renewable uncertainty comes from the DDRE-33 wind and solar scenarios (Scientific, 2025), rescaled onto transmission buses. Before any grid experiment, we run the projection-free COCO variant on the RIPE Atlas routing benchmark from the ICML 2026 paper to confirm it does not lose to the projected version on that paper's own ground.
Metrics are the regional consumption-based reduction, system emission change, $\lambda_r$, cumulative budget and leakage violation, dispatch cost, per-interval wall-clock time, and the Tapio state under all three accountings.
The decisive check is a single comparison: $\lambda_r$ for greedy average-rate shifting versus our guarded controller on transmission-congested California hours. If average-rate control shows $\lambda_r$ near zero, the core claim is wrong and the project reduces to an engineering exercise. If $\lambda_r$ is large and the guard removes most of it at small cost, the second check follows. That check asks how many region-years flip from "strong decoupling" under consumption accounting to a weaker state once displaced emissions are counted. Finally, the projection ablation and the marginal-rate arm determine which of the two methodological ingredients actually earns its place.
References
- Carbon Emission Flow in Networks
- A Power System Optimal Dispatch Strategy Considering the Flow of Carbon Emissions and Large Consumers
- Carbon Emission Flow Tracing: Fast Algorithm and California Grid Study
- Simulation and implementation strategies for optimal utilization of agricultural and forestry biomass resources based on multi-objective synergy: A case study of Fujian Province
- Hierarchical Low-Carbon Economic Dispatch with Source-Load Bilateral Carbon-Trading Based on Aumann–Shapley Method
- Optimal Anytime Algorithms for Online Convex Optimization with Adversarial Constraints
- Projection-free Online Learning over Strongly Convex Sets
- Projection-free Online Learning in Dynamic Environments
- Who affects carbon emissions? Drivers and decoupling effects of agricultural carbon emissions—evidence from Sichuan, China
- Research on Energy-Related CO2 Emissions Characteristics, Decoupling Relationship and LMDI Factor Decomposition in Qinghai
- A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network