Electricity price forecasts are no longer just read; they are acted on. Battery fleets, electric-vehicle charging operators and demand-response aggregators shift large, indivisible blocks of power into the hours a forecast says will be cheap. When such a fleet is big enough to move the price, a forecast that was accurate before anyone acted becomes wrong after everyone acts. The "cheap" hour fills up, a transmission line hits its limit, and the local price jumps. Today's forecasters are trained and scored as if the price were an external signal. This proposal treats the forecast–act–clear loop as an atomic congestion game: a game in which each player is large enough that its own move changes the cost everyone faces in discrete steps rather than smooth increments. We ask two questions. When does this loop make realized prices unstable? And how do we build a forecaster whose published number stays correct after the fleets respond to it?
The central claim is that the danger depends on whether fleets are modeled as indivisible blocks. Two facts from the congestion-game literature carry over to the grid. First, pricing rules that are right for a smooth, fluid population fail for discrete players, because they ignore the jump a player's own move causes. Second, repeated learning-style responses by heavy players can tip a market from convergence into chaos. Both predict that a passive forecaster, however accurate on historical data, will see its error on realized prices grow sharply once fleets pass a size threshold. A forecaster that predicts the price after the discrete fleet response should not degrade in the same way. If a fluid approximation of the same response works just as well, the atomic part of the story is unnecessary, and our experiments are designed to find that out.
What is known, and the gap
The forecasting side has diagnosed a "demand blindness" in sequence models. YUI: Day-ahead Electricity Price Forecasting Using Invariance Simplified Supply and Demand Curve shows that transformers such as PatchTST and iTransformer sometimes forecast lower prices when demand is doubled. Its fix, CoPiLinear, models price as a piecewise-linear function of the load rate (demand divided by available capacity). Regression Models Meet Foundation Models introduces FutureBoosting, which feeds driver forecasts from a frozen time-series foundation model into a gradient-boosted tree regressor. Both papers motivate this proposal, since they establish that structure beats raw sequence modeling. Both also leave the same hole: demand is treated as given, never as something that reacts to the price forecast itself. We call this being "response-blind," the next step after demand blindness.
Price-Aware Deep Learning for Electricity Markets supplies our main method ingredient. Its DeepWP+ model embeds a differentiable DC optimal power flow layer (the linearized market-clearing problem that produces nodal, or locational marginal, prices) and trains on nodal price error. It shows that transmission congestion amplifies small injection errors into large price swings. Its market participants, however, are passive. Here we put strategic, discrete fleets inside the clearing loop.
On the game side, Opportunities for Price Manipulation by Aggregators in Electricity Markets shows that nodal prices form a staircase in an aggregator's injection. A 0.15 MW withholding on the IEEE 6-bus system binds a line and raises profit by $2,172/h. That paper studies a single, fully informed aggregator and has no forecasting component. It tells us the response function we must forecast through is discontinuous. Optimal Taxes in Atomic Congestion Games proves that marginal-cost pricing can be worse than no intervention when players are atomic, because it ignores the shift a player's own move causes. We borrow its key insight: a price signal is only useful to a discrete player if it accounts for that player's own jump. We apply it to forecasts rather than tolls. Follow-the-Regularized-Leader Routes to Chaos in Routing Games proves that learning dynamics in congestion games become Li–Yorke chaotic once "aggression" (demand times learning rate) passes a threshold in asymmetric games, while time averages still converge. Those proofs cover only two-strategy parallel links, and extending them to general networks is listed as open. Our setting is a natural, physically grounded case beyond two links.
Buy or Sell? Energy Sharing of Prosumers on Constrained Networks and Smart Routing of Electric Vehicles for Load Balancing in Smart Grids are game models on grids. Both assume continuous bids or perfectly known price functions and include no forecaster. No paper in our corpus closes the loop from published forecast to discrete fleet response to clearing price to realized forecast error.
Core Claim
Let $d_t$ be exogenous demand at time $t$. Let $P(\cdot)$ be the DC power flow clearing map from net nodal injections to nodal prices. Let $X(\hat p)$ be the joint discrete schedule that $N$ atomic fleets choose in response to a published forecast $\hat p$. Realized prices are then $p = P(d + X(\hat p))$.
A passive forecaster targets $P(d)$. We call a forecast response-consistent if it satisfies
$$\hat p^{\star} = P\big(d + X(\hat p^{\star})\big),$$
that is, if it is still correct after fleets act on it.
When forecasters are retrained daily on realized prices, the system follows the map $\hat p_{k+1} = P(d + X(\hat p_k))$. Because $P$ is a staircase and $X$ moves in blocks, this map is not smooth. Our hypothesis is that it shows the same threshold behavior as the FoReL routing map. Below a critical value of fleet size times responsiveness, it settles. Above that value, it oscillates with period two or longer, or behaves chaotically. Day-to-day realized error then stays large even though averages over many days look fine.
Research Program
We will build a closed-loop simulator. Its base demand and renewable traces come from real data: Shanxi 15-minute series, used in both the YUI and FutureBoosting papers, and the France and Germany subsets of Real-E. These traces are mapped onto the IEEE 118-bus system used by DeepWP+, with the IEEE 6- and 14-bus systems from the aggregator paper serving as interpretable small cases.
At chosen buses we place $N$ battery or flexible-load fleets. Each fleet schedules in fixed megawatt blocks to maximize arbitrage profit against the published forecast. Each fleet's best response is a small mixed-integer program, and fleets can update either by full best response or by a damped, FoReL-style update with a tunable learning rate. Fleet share of peak load and update aggressiveness are the two axes we sweep.
The first deliverable is empirical. We will map where in this two-axis space passive forecasts become self-defeating. Diagnostics are a period detector on day-to-day realized prices and an estimate of the largest Lyapunov exponent (the average rate at which nearby trajectories separate). Because realized prices come from the clearing map, they inherit its staircase. As a theory companion, we will derive the bifurcation threshold for the 2-bus, one-line case, mapping it onto the two-strategy setting of the chaos paper. We will then check whether a threshold computed from fleet share times responsiveness predicts the empirical onset on 118 buses.
The second deliverable is the response-consistent forecaster. It keeps a strong passive model, CoPiLinear or FutureBoosting, as the predictor of $P(d)$. It then solves the fixed-point equation above with an explicit atomic response model. When evaluating whether a fleet moves a block, this response model charges the fleet the price after its own jump, which is the Chandan et al. insight applied to forecasting. A fixed point may fail to exist or may be one of several. When that happens, the forecaster publishes the time-averaged distribution of the clearing price, which the Cesàro result suggests stays stable even when single days are not. The fleet response model is fit from observed fleet behavior in the simulator, not given as an oracle, so a separate experiment reports how errors in the response model affect the result.
Baselines
The naive reference is a seasonal persistence forecast, the price at the same interval one day earlier, fed into the identical loop. It is the "do nothing" point that any learned forecaster must beat. It is also informative because a forecast that ignores recent prices partially decouples from the feedback loop.
The strongest passive aggregate-price model is CoPiLinear from YUI. On Shanxi it reports a 22.9% MAE and 38.0% sMAPE reduction over the next-best model, and over 50% of theoretical maximum storage revenue versus 40–44% for PatchTST, iTransformer, SARIMA and XGBoost. Its structure is load-rate based, so it is the hardest competitor on spikes.
FutureBoosting from Regression Models Meet Foundation Models is the strongest general-purpose pipeline. It reports up to 32.4% day-ahead and 24.9% real-time MAE reduction over zero-shot foundation models. The file gives relative improvements only; we found no reported absolute MAE values on Lacuna for either model on Shanxi.
At the nodal level, DeepWP+ from Price-Aware Deep Learning is the strongest competitor. It reports 0.6–24.7% lower average nodal price error and a 27.2% better cross-bus fairness bound on IEEE 118-bus.
Each baseline is first reproduced in open loop on the original data, to confirm that our implementation roughly matches its reported relative gains. It is then retrained daily on realized closed-loop prices, with the same features, same history, same hyperparameter search budget and the same single-GPU compute.
Two ablations isolate the key ingredient. The first is a fluid-response forecaster, identical to ours except that the fleet response is a continuous, nonatomic best response, so it is response-aware but blind to the jump each fleet's own block causes. The second is a response-blind forecaster, which runs our pipeline with the fleet model removed.
Beating the baselines means two things. In the high-share regime, we need lower realized MAE and lower conditional value at risk (the average of the worst-tail errors) of nodal price error at congested buses. And the advantage must grow with fleet share, while staying equal in the low-share regime where all methods should tie. Fleet arbitrage revenue and total system cost are reported alongside error, so that a forecaster cannot "win" by publishing numbers that freeze fleets.
Why not just…?
Why not just retrain the forecaster on realized prices every day? This is the most natural fix, and it is a baseline here, not a strawman. The concern is that daily retraining is the iterated map above. Iterated responses of this kind converge when the response is smooth and not too sensitive. The staircase evidence from the aggregator paper and the chaos results from the routing paper suggest the atomic grid case violates exactly those conditions above a threshold. If retrained baselines track the response-consistent forecaster across the whole sweep, the instability hypothesis is false, and we will report it.
Why not treat fleets as a fluid and add expected flexible load to the load rate? This is a real threat. If fleets are many and small, a fluid model should suffice, and the convergence results for large atomic games in our corpus suggest atomic outcomes approach fluid ones as players multiply. The fluid-response ablation tests this directly. We expect parity when there are many small fleets and an atomic advantage only with few large fleets near line limits. If parity holds throughout, the atomic ingredient adds nothing and the contribution reduces to response-awareness.
Why not just add fleet state of charge and size as features to FutureBoosting? Features let a regressor learn the historical average effect of fleets. They cannot capture how fleets react to the forecast now being published, because that response changes whenever the forecaster changes. We test this with a feature-augmented FutureBoosting variant under the same loop. A distribution-shift probe changes fleet size after training; a feature-based model should degrade under this shift where a structural response model should not.
Why not let the operator cap fleet ramping instead of fixing the forecast? Damping fleet updates is a legitimate market-design response, and it may remove the instability more cheaply than any forecaster. The aggressiveness axis of our sweep is exactly this lever. We will report how much damping removes oscillations, how much arbitrage value and system benefit that damping sacrifices, and whether the response-consistent forecaster reaches the same stability without the sacrifice.
Proposed Experiments
There are four decisive checks.
First, a phase diagram over fleet share and aggressiveness on IEEE 118-bus driven by Shanxi traces. The prediction is a sharp boundary beyond which passive and retrained forecasters show periodic or chaotic realized prices, and realized MAE jumps by a large factor, while the response-consistent forecaster's error stays flat.
Second, a 2-bus check comparing the analytically derived threshold with the simulated onset. Agreement would give a validated extension of the routing-chaos phenomenon to a physical network.
Third, the atomic-versus-fluid ablation, stratified by number of fleets at fixed total share. This settles whether atomicity matters.
Fourth, robustness to a misspecified fleet response model, obtained by perturbing block sizes and objectives. This shows how much response modeling error the method tolerates before it falls back to the retrained baseline.
Cross-dataset replication on the Real-E France and Germany series tests whether the boundary depends on the price regime. All closed-loop results are clearly simulation evidence. Real markets cannot be intervened on, and the claim is limited to what the grounded simulator supports.
References
- YUI: Day-ahead Electricity Price Forecasting Using Invariance Simplified Supply and Demand Curve
- Regression Models Meet Foundation Models: A Hybrid-AI Approach to Practical Electricity Price Forecasting
- Price-Aware Deep Learning for Electricity Markets
- Opportunities for Price Manipulation by Aggregators in Electricity Markets
- Optimal Taxes in Atomic Congestion Games
- Follow-the-Regularized-Leader Routes to Chaos in Routing Games
- Buy or Sell? Energy Sharing of Prosumers on Constrained Networks
- Smart routing of electric vehicles for load balancing in smart grids
- Real-E: A Foundation Benchmark for Advancing Robust and Generalizable Electricity Forecasting