🤖 machine learning

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

The paper proposes UniWind, a unified day-ahead wind power forecasting model that integrates physics-informed priors with a latent state encoder to disentangle meteorological conditions from operational states, thereby achieving high accuracy and robustness across diverse wind farms.

Original authors: Ronghui Xu, Tongxin Wu, Guozhen Zhang, Yihan Li, Chenjuan Guo, Bin Yang, Yong Li

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ronghui Xu, Tongxin Wu, Guozhen Zhang, Yihan Li, Chenjuan Guo, Bin Yang, Yong Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how much electricity a wind farm will generate tomorrow. This is a tricky job because the wind is fickle, and the turbines themselves have hidden moods.

The paper introduces a new AI model called UniWind to solve this problem. Here is how it works, explained through simple analogies:

The Problem: The "Confused" Forecast

Currently, forecasters face two main problems:

The Physics Models are too rigid: Imagine a textbook formula that says, "If the wind blows at 10 mph, the turbine makes exactly 100 watts." This works in a perfect world, but real wind farms have hills, trees, and local quirks that mess up the math. Also, the formula doesn't know if the turbine is broken or turned off by the operator.
The Data Models are too messy: Imagine a student who memorizes every past test score. If the wind was strong yesterday and the power was high, the student assumes it will be high today. But if the wind is weak tomorrow, the student gets it wrong. Worse, if the turbine was turned off (curtailed) yesterday for no reason, the student thinks "Oh, low wind means low power," confusing a mechanical shutdown with actual weather.

The Solution: UniWind's "Three-Step Recipe"

UniWind acts like a smart chef who separates the ingredients (the weather) from the cooking process (the turbine's state).

Step 1: The "Ideal Potential" (Physical Prior Estimator)

First, UniWind asks: "If this wind farm were brand new, perfectly tuned, and running at 100% efficiency, how much power could it make with this specific wind?"

The Analogy: Think of this as calculating the theoretical maximum speed of a car based on its engine and the road conditions.
The Trick: UniWind doesn't just use a generic formula. It learns a "custom fit" for each specific wind farm (like adjusting the engine for a specific driver) but keeps it grounded in the laws of physics so it doesn't make impossible predictions. It creates a "soft ceiling" or an upper limit on what is physically possible.

Step 2: The "Detective" (Latent State Encoder)

Next, UniWind looks at the difference between what should have happened (Step 1) and what actually happened in the past.

The Analogy: Imagine you predicted a car would go 60 mph, but it only went 30 mph. The detective asks: "Was the driver sleepy? Was there a traffic jam? Did the brakes get stuck?"
The Trick: UniWind analyzes these "gaps" to figure out the hidden state of the wind farm. Is the turbine in Regular Operation (running normally)? Is it Curtailment (being told to slow down by the grid)? Or is it Shutdown (broken or stopped)? It learns to recognize these patterns even though it can't "see" the switch being flipped.

Step 3: The "Smart Adjuster" (State-aware Power Corrector)

Finally, UniWind combines the "Ideal Potential" with the "Detective's findings" to make the final prediction.

The Analogy: Now that the chef knows the ingredients (weather) and the kitchen's current mood (is the oven broken? is the chef tired?), they adjust the final dish.
- If the detective says "Shutdown," the model predicts zero power, regardless of how strong the wind is.
- If the detective says "Curtailment," the model predicts power but caps it at a lower limit.
- If it's "Regular," it lets the physics take the lead.
The Result: Instead of guessing blindly, UniWind routes the prediction through different "experts" (one for shutdowns, one for normal running, one for curtailment) to get the most accurate number.

Why It Works (The Results)

The authors tested UniWind on over 20 real wind farms in China and the UK.

The "Full Shot" Test: When the model was trained on a specific farm, it beat all other methods (both physics-based and data-based) by a significant margin. It was the most accurate at predicting the next day's power.
The "Zero Shot" Test: This is the real magic. They tested the model on wind farms it had never seen before. Usually, AI models fail here because they don't know the new farm's quirks. UniWind, however, used its understanding of physics and hidden states to adapt quickly, still outperforming massive, pre-trained "foundation" models.

In a Nutshell

UniWind is like a weather forecaster who doesn't just look at the clouds. They also understand the laws of physics (how wind turns blades) and the human/machine behavior (when turbines get turned off). By separating the "weather potential" from the "operational reality," it gives a much clearer picture of how much electricity will actually be produced tomorrow.

Technical Summary: UniWind

Problem Statement

Day-ahead wind power forecasting is critical for cost-effective power-system operation, as electricity market decisions must be made prior to power delivery. Unlike conventional time-series forecasting, this task relies heavily on future meteorological conditions rather than historical power patterns. However, observed wind-farm power is an entangled signal comprising physically available power, local environmental effects, and latent operational states (e.g., shutdowns and curtailment).

Existing approaches face two primary limitations:

Physical Models: While they provide reliable constraints via physics-based inductive biases, they struggle to adapt to site-specific endogenous factors (e.g., terrain-induced circulation) and latent operational states, leading to poor generalization across different wind farms.
Data-Driven Models: These methods capture rich empirical correlations but often conflate meteorological effects with state-induced deviations. Consequently, they lack robustness under diverse operational conditions and fail to distinguish whether power deviations arise from weather changes or hidden operational interventions.

The core challenge is to construct a model that explicitly separates the dependency from meteorological conditions to physical power potential, and further to realized power under hidden-state interventions, while balancing universal physical principles with site-specific adaptation.

Methodology: UniWind

UniWind is a physics-informed state-routing model designed to decouple physically available power from operationally induced deviations. The framework consists of three sequential stages:

1. Physical Prior Estimator

This module constructs a site-calibrated physical prior representing the ideal generation potential under given meteorological conditions.

Physical Normalization: It applies a density-aware wind-speed adjustment ( $\tilde{v} = v(\rho/\rho_0)^{1/3}$ ) to align with the wind-energy relation $P \propto \rho v^3$ .
Site-Conditioned Monotone Warp: To adapt to specific sites without violating physical laws, the model uses a learnable monotonic warp function parameterized by site features ( $c$ ). This maps equivalent wind speed to effective wind speed, preserving the ordering of wind speeds.
Shared Physical Power Curve: A learnable, smooth curve captures the common physical structure of wind generation (near-zero before cut-in, nonlinear growth, and saturation). This curve is normalized by site-conditioned anchors (cut-in-like and rated-speed-like).
Upper-Bound Constraint: The prior is shaped as a soft upper envelope of available generation. A loss function ensures the prior covers observed power while remaining tight during regular operation, preventing overestimation.

2. Latent State Encoder

This module infers unobservable, time-varying operational states by analyzing discrepancies between observed power and the physical prior.

Historical State Encoding: It constructs state tokens from historical power, the physical prior, residuals, and their temporal differences. These are encoded into historical state embeddings.
Future State Retrieval: Using a mechanism similar to attention, the model treats future weather embeddings as queries and historical weather/state embeddings as keys and values. This retrieves future latent state embeddings by transferring patterns from similar historical meteorological contexts.
Refining: A causal Temporal Convolutional Network (TCN) refines the combined state sequence, fusing it with weather features to produce final state embeddings.

3. State-Aware Power Corrector

This stage transforms the physical prior into realized power forecasts by applying state-specific corrections.

Site-Context Conditioning: Site features and historical power statistics are encoded and injected into the state embeddings to account for persistent site-specific characteristics.
Supervised State Routing: A router assigns each time step to one of three operating-state experts: Regular, Curtailment, and Shutdown. The routing is supervised by state-prior labels derived from wind-power behavior (not ground-truth operational annotations).
Bounded Expert Correction: Each expert applies a distinct correction mechanism:
- Shutdown: Sets output to zero.
- Curtailment: Learns an upper-bound ratio.
- Regular: Learns a multiplicative correction under the rated-capacity bound.
  The final forecast is a routing-weighted aggregation of these bounded expert outputs.

Key Contributions

Unified Framework: Proposes UniWind, the first model to explicitly factorize day-ahead wind power forecasting into meteorology-driven physical potential and state-dependent realized generation.
Site-Calibrated Physical Prior: Develops a Physical Prior Estimator that combines shared wind-energy structures with site-conditioned monotone calibration and physical upper-bound regularization, avoiding the rigidity of fixed formulas and the overfitting of unconstrained data-driven methods.
Latent State Modeling: Introduces a Latent State Encoder and State-aware Power Corrector to model unobservable operational states. The system retrieves future state representations from historical weather-state patterns and applies knowledge-supervised, bounded corrections.
Empirical Validation: Conducts extensive experiments on over 20 real-world wind-farm datasets, demonstrating superior accuracy and robustness in both full-shot and cross-farm zero-shot settings.

Experimental Results

The authors evaluated UniWind on 24 real-world datasets from the UK and three Chinese provinces (Shandong, Shanxi, Anhui), using 15-minute interval data.

Full-Shot Performance: UniWind achieved the best performance across all representative datasets and metrics (MAE and RMSE) compared to physical models, statistical baselines (LightGBM, XGBoost), time-series models (PatchTST, iTransformer), and renewable-energy specific models. For instance, on the SD_A dataset, UniWind reduced MAE to 12.05 compared to 14.90 for LightGBM and 28.64 for PowerCurve.
Zero-Shot Performance: In cross-farm zero-shot settings (training on all datasets except the target), UniWind substantially outperformed wind-specific foundation models (WindFM) and general time-series foundation models (Chronos-2, Moirai). This indicates that combining future NWP covariates with transferable physical priors is more effective than domain-specific or generic pretraining alone.
Robustness: Under varied conditions (low/high wind, regular/abnormal states), UniWind consistently outperformed baselines, particularly in "abnormal" states where other models struggled to distinguish between meteorological changes and operational interventions.

Significance and Claims

The paper claims that UniWind addresses a critical gap in wind power forecasting by explicitly separating physical generation potential from state-dependent realized power. By integrating physics-informed constraints with latent state routing, the model avoids the pitfalls of fixed physical formulas (which lack adaptability) and pure data-driven approaches (which conflate physical and operational effects). The authors assert that this approach enables more accurate and robust day-ahead forecasting, which is essential for reliable power-system operation and electricity market decision-making. The results suggest that explicitly modeling the dependency chain from meteorology to physical potential to state-mediated realization is a necessary step for advancing wind power prediction beyond current state-of-the-art methods.