UniWind: Day-Ahead Wind Power Forecasting via Physics-Informed State Routing

arXiv:2607.01670 · cs.LG · Submitted 2026-07-02 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "UniWind: Day-Ahead Wind Power Forecasting via Physics-Informed State Routing".

Jane: The gist: UniWind proposes a physics-informed state-routing model for day-ahead wind power forecasting that explicitly factorizes the forecasting process into meteorology-driven physical power potential and state-dependent realized generation.

Tom: First, who's behind it and why it matters.

Title and authors: Jane: So, let's talk about what this paper is actually proposing in terms of its core idea. The title says "Physics-Informed State Routing," which sounds pretty technical.

Tom: It means they are using the physics of wind power—how wind speed relates to power—to build a starting point, and then routing that through a system that handles those unobservable operational states.

Lu: They construct a site-calibrated physical prior, which is essentially an upper bound on the available generation under ideal conditions, by applying things like density-aware wind-speed adjustments and some kind of monotone warp map to get an effective wind speed.

Meng: So they are building a constraint based on physical laws, but that sounds like a big assumption. What if the physical model itself is slightly off for this specific turbine site?

Tom: They address that by making it site-calibrated, meaning they adapt the prior using something called a function hϕ(c) in R squared to map the wind speed to an effective wind speed based on the local conditions.

Lalam: And they use this physical upper envelope constraint to shape a soft upper envelope of available generation through a loss function called Lprior, which acts as their physical prior regularization loss.

The paper's summary: Jane: So, if we boil down the summary, UniWind first figures out the maximum power physically possible based on wind speed and site conditions, and then it looks at historical data to build tokens that describe the operational states.

Tom: They use discrepancies between observed power and that physical prior to construct discrepancy-aware historical state tokens s h, which they then encode into D-dimensional embeddings.

Lu: And they also have a weather encoding process involving both a per-step MLP encoder and a temporal weather encoder to get context-aware weather embeddings h w.

Meng: That sounds like a lot of layers just to describe the environment. How do you actually use those historical state tokens and the future weather embeddings together?

Lalam: Future latent state retrieval is done by treating future weather embeddings as queries against historical weather embeddings and historical state embeddings to get those future latent state embeddings s f.

Jane: And finally, they have this State-aware Power Corrector that takes the final state embedding for each time step and routes it to supervised experts using a learnable router R(z˜) to produce the final forecast pˆ.

The paper's improvements: Tom: One major improvement is how they decouple the meteorological effects from the operational state deviations, which means they can follow the physical wind-to-power relationship in regular times while correcting for deviations in abnormal times.

Lu: That separation allows them to model both meteorology-driven generation potential and its state-dependent realization, which is a big step because before this, data-driven models tended to conflate those two things.

Meng: So the system can distinguish between a power drop because the wind is slow and a drop because someone intentionally cut output, that's what I'm hearing. That sounds like it solves the core problem for deployment.

Jane: They also introduce state transition prediction, where they try to infer future operational states based on historical patterns under similar meteorological conditions so they can anticipate how a weather change might trigger an operational shift.

Lalam: And the final correction step is crucial, where the State-aware Power Corrector routes each time step's embedding to supervised operating-state experts applying bounded, state-specific corrections to produce the final forecast pˆ.

Conclusion: Tom: So UniWind seems pretty strong because it maintains high accuracy across all regimes when evaluated under varied meteorological and operational conditions, even showing the lowest Mean Absolute Error in all four regimes.

Jane: The paper suggests that by separating physical power variations from state-induced deviations, you get a model that works well in regular operating conditions while still correcting those deviations when things go abnormal.

Lu: The ablation studies confirm that each component contributes to the forecasting accuracy, which means we can see exactly what part of the system is actually helping the forecast.

Meng: I'm looking at their parameter count now, and they only have four million parameters and take about zero point seven two milliseconds per sample for inference time, which suggests it could be feasible to run this on actual wind farm data.

Lalam: So in short, the UniWind: Day-Ahead Wind Power Forecasting via Physics-Informed State Routing model is a day-ahead wind power forecasting model that captures the structured dependency from meteorological conditions to physically available power and realized generation under latent operational states.

Tom: That's what we have here. It’s a framework that provides a meaningful upper-envelope reference through the physical prior, and then uses state-aware correction to translate that reference into realistic day-ahead power forecasts.

Jane: It really moves beyond just predicting what the weather says about wind speed and tries to predict what actually happens at the turbine level under hidden operational conditions.

Lu: It’s a solid piece of research because it addresses that structural dependency you mentioned earlier, making the model more adaptable across different wind farms and conditions.

Meng: For deployment, having a model that can reliably distinguish between physical limits and operational decisions is what we need to move toward real-world use cases.

Lalam: This paper on UniWind is about using physics to ground the forecast before applying AI to handle the messy reality of how power is actually generated.

East China Normal University · Southern University of Science and Technology Shenzhen · TsingRoc.ai Institute of Technology Tsinghua University

cs.LG

Submitted: 2026-07-02

Updated: 2026-10-15

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: The gist: UniWind proposes a physics-informed state-routing model for day-ahead wind power forecasting that explicitly factorizes the forecasting process into meteorology-driven physical power

Key concepts

Physical Prior Estimator: This module builds an initial forecast by modeling the physically available power. It uses a site-calibrated monotone warp and a shared physical power curve to create an upper envelope, representing the maximum possible wind generation under ideal conditions.
Latent State Encoder: This component models unobservable operational states by analyzing discrepancies between observed power and the physical prior. It encodes these historical state tokens into embeddings, allowing the model to capture how past operational conditions influence current generation.
State-aware Power Corrector: This final module takes a time step's latent state embedding and routes it to specific experts using a learnable router. It applies bounded, state-specific corrections to transform the physical prior into the final, realistic power forecast.
Physical Normalization: This step adjusts raw wind speed measurements by applying a density-aware factor ($ ilde{v} = v / ho_0^{1/3}$). This normalization aligns the input data with the fundamental wind-energy relationship where power is proportional to the cube of wind speed.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Embeddings — 1× in this paper · explained in 2 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

The gist: UniWind proposes a physics-informed state-routing model for day-ahead wind power forecasting that explicitly factorizes the forecasting process into meteorology-driven physical power potential and state-dependent realized generation.

How it works

UniWind first constructs a site-calibrated physical prior to estimate the physically available power under ideal conditions, which serves as an upper bound on realized power. This is achieved through several steps:

Physical normalization involves applying a density-aware wind-speed adjustment to obtain the equivalent wind speed v˜ = v / ρ0(1/3), following the wind-energy relation Pwind ∝ ρv3.
Site-conditioned monotone warp maps equivalent wind speed to effective wind speed using a function hϕ(c) ∈ R2, where [a, b]⊤ = hϕ(c) ∈ R2.
A physical upper-bound constraint shapes this prior into a soft upper envelope of available generation using the loss Lprior.

How it works (continued)

Second, UniWind introduces modules to model unobservable operational states and their effects on generation:

The Latent State Encoder characterizes historical operating states by using discrepancies between observed power and the physical prior to construct discrepancy-aware historical state tokens, which are then encoded into D-dimensional historical state embeddings s h.
It also encodes the full weather sequence through a process involving an MLP-based per-step weather encoder and a temporal weather encoder to obtain context-aware weather embeddings h w.
Future latent state retrieval is performed by treating future weather embeddings as queries against historical weather embeddings and historical state embeddings to retrieve future latent state embeddings s˜f.
Finally, the State-aware Power Corrector routes each time step's final state embedding to supervised operational-state experts using a learnable router R(z˜) and applies bounded, state-specific corrections to produce the final forecast pˆ.

Training Objectives

The model is trained by minimizing a comprehensive loss function L, which combines three components:

Mean Squared Error loss (Lmse) for forecasting supervision: Lmse = xˆ − x2F + yˆ − y2F.
Physical prior regularization loss (Lprior): Lprior = 1/omega Σ t∈omega h pt - p prior t + λtight / omegareg Σ t∈omegareg h p prior t - pt - ϵslacki.
Supervised state routing loss (Lrouter): Lrouter = CE(ˆl,l). Full-shot prediction results show that UniWind achieves the best full-shot performance across all representative datasets and evaluation metrics. Zero-shot prediction results also show that UniWind substantially outperforms WindFM, Chronos-2, Moirai, and Moirai-PT across all datasets and metrics.

Key Findings

UniWind remains strong across all regimes when evaluated under varied meteorological and operational conditions. This supports the core motivation of UniWind: by separating physical power variations induced by wind speed from power deviations induced by operational states, UniWind can follow the physical wind-to-power relationship in regular regimes while correcting deviations in abnormal states. The ablation studies confirm that each component contributes to the forecasting accuracy. Furthermore, UniWind has only 4.0M parameters and requires 0.72 ms per sample for inference time.

Conclusion

In this paper, we propose UniWind, a day-ahead wind power forecasting model that captures the structured dependency from meteorological conditions to physically available power and realized generation under latent operational states. The proposed site-calibrated physical prior and state-aware correction help mitigate dependence on NWP inputs. Future work will explore the integration of additional meteorological observations to better capture local weather dynamics and further improve forecasting reliability. The proposed physical prior provides a meaningful upper-envelope reference, while state-aware correction is necessary to translate this reference into realistic day-ahead power forecasts.

References

[1] Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, et al. Chronos-2: From univariate to universal forecasting. arXiv preprint arXiv:2510.

Improvements for AI systems

No longer conflating meteorological effects with state-induced deviations: UniWind explicitly decouples physically available power from operationally induced deviations, thereby modeling both meteorology-driven generation potential and its state-dependent realization. This allows the system to distinguish between a drop in power due to low wind speed (physical potential) and a drop due to intentional shutdowns or curtailment (latent state).
Site-specific physical prior construction: The model can now construct site-calibrated estimates of physically available power, providing a reliable prior for latent state modeling and state-aware correction by combining site-conditioned monotone warping with a shared physical power curve. This ensures the physical constraints are adapted to the specific wind farm's topography and turbine characteristics.
State transition prediction: The system can now infer future operational states based on context: The encoder retrieves future state patterns from historical states under similar meteorological conditions, enabling it to anticipate how a change in weather might trigger an operational shift.
State-aware correction for realism: The final forecast is enhanced by learning state-specific adjustments, as the State-aware Power Corrector routes each timestamp to supervised operating-state experts and applies bounded state-specific corrections to produce realistic forecasts under hidden operating conditions. This moves predictions beyond the ideal physical envelope to reflect actual realized power.
Robustness across operational regimes: The system maintains high accuracy in diverse conditions: UniWind achieves the lowest MAE in all four regimes (low-wind, regular state, low-wind, abnormal state, high-wind, regular state, and high-wind, abnormal state). This demonstrates that the model successfully separates physically available power variations induced by wind speed from power deviations induced by operational states.

Abstract

Accurate day-ahead wind power forecasting is essential for cost-effective power-system operation. Wind power generation is primarily driven by future meteorological conditions while retaining temporal dependencies in power generation. In practice, observed wind-farm power reflects both meteorology-driven generation potential and deviations associated with local environmental and operational factors. Existing physical models provide useful constraints but adapt poorly across wind farms, whereas data-driven models can capture rich correlations but do not explicitly structure these two sources of variation. To bridge this gap, we propose UniWind, a wind power forecasting framework that integrates physics-informed generation priors with data-driven state-aware corrections. UniWind first employs a Physical Prior Estimator to construct a site-calibrated physical prior by combining a site-conditioned monotone warp with a shared physical power curve. It further applies a physical upper-bound constraint to shape this prior as a soft envelope of available wind power generation. UniWind then uses a Latent State Encoder to construct latent power-state representations from historical prior-power discrepancies and future meteorological context, and transforms the physical prior into final power forecasts through a State-aware Power Corrector with pseudo-label-guided routing and bounded expert correction. Experiments on 24 real-world datasets demonstrate UniWind's strong forecasting accuracy, robustness, and generalization across full-shot, cross-farm zero-shot, and cross-region zero-shot settings.