Grid operators increasingly use learned load forecasters to set the prices they broadcast to automated devices such as home batteries, smart thermostats and electric-vehicle chargers. Customers and aggregators who leave a program can demand that their historical data be deleted. "Machine unlearning" means removing a subset of training data from a trained model without retraining it from scratch. Deletion makes the forecaster somewhat worse, and the usual way to judge an unlearning method is to look at that degradation in isolation: forecast error, whether a membership-inference attacker can still tell the deleted data was used, and at most the cost of dispatching generators against the forecast. Our argument is that this is the wrong yardstick when the forecaster's output becomes a price that thousands of cost-minimizing devices react to. Small forecast errors that land near a cheap price window can move many devices into the same hour together. The result is a "rebound peak," an artificial demand spike that can exceed the natural peak the price was meant to flatten. Two forecasters with similar error, similar privacy scores and even similar open-loop dispatch cost could therefore produce very different peaks once the devices respond.
We propose to measure this directly. We will build an assessment that scores each unlearning method by its effect on the whole closed loop: forecaster, then price signal, then device responses, then feeder load. We will then test whether that score ranks unlearning methods differently from open-loop metrics. The forecaster is trained across aggregators with a privacy-preserving distributed algorithm, so we can also ask whether the accuracy lost to privacy noise and the accuracy lost to deletion add up to synchronized peaks. Finally, we will test whether a simple response-aware change to an existing task-aware unlearning method closes any gap we find. The two-sentence version is this: unlearning a forecaster that sets prices should be judged by the demand peaks the resulting prices cause, not only by forecast error. We will measure whether those two judgments disagree, and fix the unlearning objective if they do.
Core Claim
The claim is falsifiable. When a forecaster sets prices for responsive automated agents, rankings of unlearning methods built from open-loop metrics will disagree substantially with rankings built from closed-loop peak and feasibility metrics. Open-loop metrics here include forecast error, membership-inference forgetting and operational dispatch cost. Closed-loop metrics include peak-to-average ratio, the size of capacity exceedances and rebound-peak magnitude.
We expect the disagreement because agent response is sharply nonlinear in the timing of price changes. Devices pile up at the start of cheap billing windows, as Demand Side Management in the Smart Grid: an Efficiency and Fairness Tradeoff documents with Pecan Street households. Even a modest fraction of highly responsive agents can destabilize aggregate demand, as shown in Roozbehani et al. and Gupta et al.. An unlearning method that spreads its error evenly over the day should therefore be far less harmful than one that shifts the forecast of when the peak occurs, even if both have the same mean squared error.
This claim extends the finding that motivates the project, Task-Aware Machine Unlearning and Its Application in Load Forecasting (TAMU). TAMU showed on the Texas backbone system that forecast error barely correlates with generator cost ($r = 0.073$). It also showed that deleting 20% of training data by "complete" unlearning raised generator cost by 20%. It responded by pushing dispatch-cost gradients into the unlearning update through a three-level optimization. TAMU treats load as passive: the forecast feeds a centralized dispatch and nothing reacts to it. We keep TAMU's insight that operational metrics matter, but add the feedback loop through device behavior that TAMU leaves out. We ask whether dispatch cost is itself a sufficient operational metric once demand responds to the forecast.
Research Program
Simulated closed loop. We first build a reproducible closed-loop simulator out of pieces already in the literature. A neural day-ahead load forecaster is trained on household traces grouped into aggregators. Its forecast determines a broadcast price under one of three coordination regimes, chosen to span how much protection the market design already gives:
- Static peak/off-peak tariff. The regime known to cause boundary synchronization (Jacquot et al.).
- Adaptive pricing with a damping fee. The fee penalizes day-to-day changes in each agent's storage schedule, from Decentralised Control of Micro-Storage in the Smart Grid.
- Dual-descent pricing with physical feedback. Prices are updated from measured aggregate power, with step sizes chosen so demand stays under capacity, from Distributed Resource Allocation Using One-Way Communication with Applications to Power Networks.
In all three regimes the forecaster sets the starting price and the day-ahead schedule. Agents are heterogeneous batteries and electric-vehicle chargers, and each solves a small convex cost-minimization problem. Feeder limits come from the IEEE-33 bus network with renewable scenarios from A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network.
Privacy constraint. The forecaster is trained across aggregators with a privacy-preserving distributed scheme. Each aggregator keeps its raw traces local, and a gradient-based variant of the alternating direction method of multipliers adds noise only to the primal updates. That design is chosen because Privacy Amplification by Iteration for ADMM with (Strongly) Convex Objective Functions shows such noise spreads into the dual variables and strengthens the privacy guarantee over iterations.
This choice is constrained by Differentially Private Distributed Convex Optimization via Functional Perturbation. That paper proves that noise added to exchanged messages must either fail to give differential privacy (if it vanishes) or leave a permanent accuracy floor (if it does not). We take this seriously: we treat the privacy noise level as an experimental factor that sets the accuracy floor, and we run a functional-perturbation variant at small scale for comparison.
Distributed training also changes what retraining costs. Retraining after a deletion means rerunning the whole multi-party protocol and spending more privacy budget, not only more compute. This is why approximate unlearning, including the decentralized certified variant of Certified Unlearning in Decentralized Federated Learning, is worth studying here even though the forecaster itself is small.
Assessment. The assessment follows the multi-axis logic of MUSE and Deep Unlearn, which showed that single metrics let unlearning methods "cheat," for example by collapsing the model. Each unlearning event is scored on five groups of measures:
- Forgetting: membership-inference distinguishability of the withdrawn aggregator's traces, plus re-learning speed following Golatkar et al..
- Fidelity: agreement with the forecaster retrained without that aggregator.
- Open-loop utility: forecast error and dispatch cost.
- Closed-loop stability: peak-to-average ratio, maximum and integrated exceedance of feeder capacity, and rebound-peak height relative to the natural peak.
- Fairness and sustainability: the billing fairness index of Jacquot et al., and behavior under a sequence of up to ten withdrawals. The sequence tests MUSE's sustainability axis, which TAMU explicitly leaves open.
The central analysis compares method rankings across these groups. We compute rank correlation between the open-loop and closed-loop orderings across methods, deletion sizes, privacy levels and pricing regimes. We also locate where in the daily load curve the unlearning-induced error falls.
Response-aware unlearning. If closed-loop and open-loop rankings disagree, we test a direct remedy. We extend TAMU's differentiable optimization layers with one more layer: a smoothed version of the agents' collective response. Each agent's cost-minimization is regularized with an entropy term so that the aggregate load is a differentiable function of price. The unlearning update then minimizes
$$\mathcal{L}_{\text{forget}}(\theta) + \lambda\, C_{\text{dispatch}}(\theta) + \mu\, \max_t \big(\textstyle\sum_i q_i(p_\theta, t) - Q\big)_+ ,$$
where the symbols mean:
- $\theta$ are the forecaster weights.
- $\mathcal{L}_{\text{forget}}$ is the unlearning loss on the withdrawn data.
- $C_{\text{dispatch}}$ is TAMU's operational cost.
- $q_i(p_\theta, t)$ is agent $i$'s smoothed demand at hour $t$ under the price $p_\theta$ that the forecaster induces.
- $Q$ is feeder capacity.
- $\lambda$ and $\mu$ are weights.
The added ingredient is the last term, which penalizes demand above capacity in the worst hour once agents respond.
Baselines
Do nothing and retrain. The naive reference point is to keep the original forecaster, which forgets nothing. Its closed-loop peaks are the floor for every deletion. The gold standard is exact retraining without the withdrawn aggregator, run through the same private distributed protocol at the same privacy level. Retraining matters for a second reason: it shows how much of any rebound peak is caused by losing the data itself rather than by approximate unlearning.
Generic unlearners. We include gradient ascent and its regularized variant NG+ as the weak generic baselines. Deep Unlearn found gradient ascent poor at retaining utility, and MUSE found that gradient-ascent-style methods either under-unlearn or over-unlearn in ways membership attacks detect. We also include Masked Small Gradients, the strongest method in Deep Unlearn that carries over to a forecasting architecture. Deep Unlearn reports it as up to 5.3x faster than retraining. Convolution Transpose (10.59x) is included only if the forecaster is convolutional.
Strongest competitor: TAMU. TAMU is the strongest existing method because it is the only one that already optimizes an operational grid objective during unlearning. On the Texas backbone system it reports the $r = 0.073$ correlation between forecast error and generator cost and the 20% cost increase under complete unlearning that TAMU's correction targets. We found no reported result on Lacuna for TAMU's effect on peak-to-average ratio, capacity exceedance or any closed-loop demand-response metric.
Decentralized certified unlearning. The method of Certified Unlearning in Decentralized Federated Learning is included as the decentralized competitor.
Fairness of comparison. Every method uses the same forecaster architecture, the same training data and aggregator split, and the same deletion requests. Every method gets the same hyperparameter search budget, tuned on a validation week against its own objective, and the same wall-clock limit set at a fixed fraction of retraining time.
Ablation. The key ablation is response-aware unlearning with $\mu = 0$. This is essentially TAMU run on our distribution network, and it isolates whether the agent-response term does any work.
What counts as beating TAMU. Response-aware unlearning must reduce maximum capacity exceedance and rebound-peak height under the static-tariff and damping-fee regimes. It must do so while staying within TAMU's dispatch cost and membership-inference distinguishability, using confidence intervals over seeds and test weeks.
Why not just…?
Why not just use TAMU's dispatch cost as the operational metric? This is the most direct threat. If dispatch cost already tracks peak harm, the closed-loop assessment adds nothing. TAMU's own evidence shows how badly statistical and operational metrics can diverge, but dispatch cost is computed against passive load and has no channel through which agents react to a mispriced hour. The rank-correlation analysis settles this question directly. If dispatch-cost rankings and closed-loop rankings agree (Kendall's tau above roughly 0.8 in every regime), we will report that the closed loop is unnecessary and that TAMU suffices.
Why not just retrain from scratch, since load forecasters are small? For a centralized forecaster this is a fair objection, and compute is not our argument. In the distributed private setting, retraining reruns a multi-party protocol and spends more privacy budget for every withdrawal, and withdrawals arrive continually. More importantly, retraining is included as a baseline precisely because it also degrades accuracy. If retrained forecasters produce rebound peaks as large as approximate ones, then the problem lies in how deletions are absorbed by the pricing loop and not in approximate unlearning. The assessment is designed to reveal that either way.
Why not just rely on a desynchronizing market design such as the damping fee or feedback dual descent? This is a genuine threat. Voice et al. report that the damping fee cuts the root-mean-squared gap between predicted and actual demand from 31.5 to 14.1 and cuts wholesale cost volatility from 6.6% to 5.3%. They also report recovery after half the batteries fail. Magnússon et al. guarantee capacity is never exceeded during their price iterations. If these mechanisms absorb unlearning-induced error, unlearning quality is irrelevant to stability under them, and that is precisely why they are two of our three regimes. We expect the feedback regime to be the most robust, but its guarantee depends on a sufficiently high starting price and a bounded step size, both set from the forecast. It is untested when forecasts or utilities carry noise, a tension the research file identifies between functional perturbation and feasibility preservation. A null result under these two regimes would itself be useful, because it would tell regulators which market designs make data deletion safe.
Proposed Experiments
Data. Household traces come from the Pecan Street dataset used by Jacquot et al., grouped into 10 to 30 aggregators, with electric-vehicle and heating flexibility as in that paper. The damping-fee regime is checked against UK national load and APX-ENDEX/BM Reports prices. Before adding unlearning, we first reproduce two published results as calibration checks: Voice et al.'s RMS improvement and Jacquot et al.'s finding that the hourly-proportional mechanism comes within 0.08–0.1% of the social optimum. Feeder constraints use IEEE-33 bus with the DDRE-33 renewable scenarios. Dispatch cost uses IEEE 14-bus, where we first try to reproduce TAMU's 20% cost increase.
Design. Deletions remove 5%, 10% and 20% of data, either as whole aggregators or as households with the most flexible load. The second option targets the users whose withdrawal most changes the forecast near price boundaries. We run three privacy noise levels plus a non-private control, and sequences of up to ten withdrawals.
Decisive checks. There are three:
- Rank agreement. The rank correlation between open-loop and closed-loop orderings of unlearning methods across all regimes. The main claim fails if the correlation is high everywhere.
- Response-aware ablation. The $\mu$ ablation must cut rebound-peak height without worse membership-inference distinguishability. If it does not, a better unlearning objective is not the fix, and the contribution narrows to the assessment.
- Feasibility under noise. Whether Magnússon et al.'s capacity guarantee survives when the starting price comes from a privately trained, then unlearned, forecaster. We will count exceedance events in that regime; any nonzero count is direct evidence that privacy noise and unlearning together break a guarantee previously proved only for exact inputs.
References
- Demand side management in the smart grid: An efficiency and fairness tradeoff
- Volatility of Power Grids Under Real-Time Pricing
- Optimal Real-Time Coordination of Energy Storage Units As a Voltage-Constrained Game
- Task-Aware Machine Unlearning and Its Application in Load Forecasting
- Decentralised Control of Micro-Storage in the Smart Grid
- Distributed resource allocation using one-way communication with applications to power networks
- A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network
- Privacy Amplification by Iteration for ADMM with (Strongly) Convex Objective Functions
- Differentially Private Distributed Convex Optimization via Functional Perturbation
- Certified Unlearning in Decentralized Federated Learning
- MUSE: Machine Unlearning Six-Way Evaluation for Language Models
- Deep Unlearn: Benchmarking Machine Unlearning for Image Classification
- Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks