acceptodds
Under review as a conference paper at ICLR 2027

A Conditional Koopman World Model for Distribution-Aware Electricity Load Forecasting

Abstract

Electric load and price time series exhibit typical damped oscillation patterns with periodic cycles and spike mean reversion, which are dominated by the dynamic evolution of power supply and demand contradictions. Existing time series forecasting models rarely incorporate such physical inductive biases, making it challenging to simultaneously achieve stable long-horizon rollouts, event-conditioned forecasting, and counterfactual simulation. To address this issue, this paper proposes the Power World Model (PWM), a conditional Koopman world model for distribution-aware electricity forecasting. Leveraging the dual core mechanisms of amplitude-conserving rotation and amplitude-dissipating contraction inherent to damped oscillation, PWM constructs a latent state space through delay embedding and designs a structurally decomposable condition-affine Koopman operator. The operator structurally limits the spectral radius to no more than 1, ensuring bounded and stable long-horizon forecasting results. Unlike conventional conditional forecasting methods that integrate conditional information at the input layer, our method embeds conditional signals into the state transition operator, enabling efficient event-conditioned forecasting and low-cost counterfactual rollouts. Under a unified experimental protocol across multiple public datasets and a private power dataset, PWM achieves state-of-the-art point prediction accuracy, outperforming existing mainstream time series forecasting models. Mechanism validations on synthetic load-shedding events demonstrate that PWM can effectively improve prediction performance in extreme scenarios.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, Koopman, electricity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g9UvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Power World Model for electricity forecasting. This framework uses learnable dissipation and rotation components to construct a Koopman operator that captures periodic patterns and let disturbances decay gradually. Experiments on five public benchmarks and a private electricity-market dataset demonstrate competitive forecasting performance.

Critical strengths

The structure of the framework is complete .The experimental results on five public benchmarks are mostly competitive.

Critical weaknesses

Table 5 examines the stability of different transition operators, while Table 6 focuses on encoder ablations. Neither directly shows whether the Dissipation–Rotation Parameterization improves MSE or MAE over alternative transition operators. The paper also introduces Event Gating and a Mode Library, but Appendix A.2 states that the experiments use only a single conditional operator. The contribution of the proposed transition design to forecasting accuracy therefore remains unclear.

Table 3 reports improvements on synthetic events generated using the same protocol during training and testing. It is unclear whether these gains would hold for unseen event magnitudes or patterns. The experiments also compare conditional PWM only with its unconditional version. A simple input-concatenation baseline would help show whether modifying the Koopman transition is more effective than directly providing the model with event information.

The Quantile Head predicts the 0.1, 0.5, and 0.9 quantiles using standard quantile regression techniques. Both conditional and unconditional PWM use the same head, so the CRPS improvement in Table 3 does not establish its individual contribution. Without a comparison with other probabilistic prediction heads, it is difficult to assess what this component adds to the framework.

Questions for authors
  1. Table 3 compares conditional PWM only with its unconditional version under the same synthetic load-shedding protocol. How does it compare with an input-concatenation baseline? Does it generalize to event patterns that are not seen during training?
  2. Table 5 evaluates the stability of different Koopman operators, but not their forecasting accuracy. How do the unconstrained and unitary variants compare with PWM in terms of MSE and MAE under the same encoder and experiment setup?
  3. The Ethics Statement says PrivEM is not redistributed, while the Reproducibility Statement mentions a de-identified version in the supplementary material. What data are actually shared, and is this permitted by the data-use agreement?
Reviewer_Kg29OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PWM, a conditional Koopman model for electricity forecasting. It combines Transformer-based representations with a dissipation–rotation structured operator to ensure stable latent dynamics. External conditions are incorporated into the transition operator, and a quantile head supports probabilistic forecasting. Experiments demonstrate competitive performance on public and private electricity datasets.

Critical strengths
  1. The dissipation–rotation parameterization provides a meaningful inductive bias with theoretical stability guarantees.
  2. The model achieves competitive forecasting accuracy on multiple electricity datasets.
  3. Incorporating external conditions into Koopman dynamics is an interesting approach to event-aware forecasting.
Critical weaknesses
  1. The model uses fixed Koopman refinement steps followed by direct decoding, so the theoretical stability guarantee does not directly establish stable long-horizon forecasting.
  2. The contribution of the proposed operator is unclear without comparisons against simpler latent transformations.
  3. Conditioning provides inconsistent improvements, and comparisons with standard conditioning methods are missing.
  4. Counterfactual forecasting and probabilistic calibration are insufficiently evaluated. Multi-seed results are also missing.
Questions for authors
  1. How does the stability guarantee apply to long-horizon predictions when the operator is used only for fixed-step latent refinement?
  2. Can the authors compare the proposed operator with unconstrained, unitary, and simpler latent transformations?
  3. Does operator-based conditioning outperform standard input-level conditioning?
  4. Can the authors provide stronger counterfactual evaluations, probabilistic calibration results, and multi-seed experiments?
Reviewer_p9cvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work develops a conditional Koopman operator-based world model for electricity price forecasting. The developed world model involves two mechanisms: amplitude-conserving rotation and amplitude-dissipating contraction, for capturing the damped oscillation patterns within the time-series electricity prices. The results show the competitive forecasting performance against baselines.

Critical strengths

This paper aims to develop a world model for electricity price forecasting.

Critical weaknesses
  1. It lacks a clear motivation for the key damped oscillation assumption. Without theoretical or empirical evidence, the key assumption is asserted rather than evidence-grounded. Moreover, the mentioned "amplitude-preserving phase rotation" does not make sense to price forecasting in 15-minute resolution, as the prices are largely driven by supply-demand mismatch and market bids. The whole problem setting is not quite relevant.

  2. This work adopts the conditional Koopman world model for price forecasting without clearly justifying why the adopted model fits or capturing fundamentals of electricity market rules, limiting the contribution and soundness.

  3. The evaluation is very limited without comparisons with other world models or using recent real-world electricity price data.

Questions for authors

It seems that the dissipation-rotation decoupled Koopman-based method is already established in the literature, see Ref[1]. How does the proposed method differ from the existing ones? Also, the literature review does not include this existing study which is closely related to the key contribution. Ref[1]: KoopGen: Koopman generator networks for representing and predicting dynamical systems with continuous spectra.

For the public datasets used in the study, i.e., ECL, ETTh2, ETTm2, they are normally used for electricity load forecasting, not price forecasting. Forecasting the electricity price is more challenging than the load forecasting, since the price is not only affected by the grid loads, but also other factors.

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.