Short-Term Forecasting of Heterogeneous Wind Power Fleets: Global Temporal Fusion Transformer Leveraging Static Covariates

Viktor Walter; Andreas Wagner

cs.LGNov 4, 2026

Short-Term Forecasting of Heterogeneous Wind Power Fleets: Global Temporal Fusion Transformer Leveraging Static Covariates

Authors: Viktor Walter, Andreas Wagner

Organizations: Faculty of Management Science and Engineering, Karlsruhe University of Applied Sciences, Karlsruhe, Germany · Fraunhofer Institute for Industrial Mathematics, Kaiserslautern, Germany

Abstract

Accurate short-term wind power forecasting is critical for grid stability and electricity trading. While transformer-based deep learning is state of the art, industry practice relies mostly on local models trained for individual assets. This study investigates whether a global temporal fusion transformer (TFT), trained jointly on a heterogeneous portfolio of wind farms across Germany, outperforms site-specific local models in 48-hour-ahead forecasting. Using numerical weather prediction (NWP) data from the regional ICON-D2 and the global ECMWF HRES models, we compare the global TFT against local models and a linear regression baseline. The global model leverages static covariates, such as farm age, altitude above sea level, hub height, rotor diameter and power curve parameters, to learn shared representations across sites. We evaluate on 100 synthetic wind farm profiles and a real portfolio of 11 German wind farms, across two seasonal test periods with intermediate retraining. The global TFT outperforms the local models for 87% of the synthetic and 77% of the real wind farms and clearly surpasses the linear baseline. Wilcoxon signed-rank tests confirm that these gains are significant across both seasons and datasets. Crucially, this superiority depends on static covariates: without asset-specific context, the global model fails to generalize. The TFT's variable selection network (VSN) reveals the learned feature selection. Altitude emerges as a relevant static discriminator on both datasets, while hub height is decisive on the real portfolio. For dynamic inputs, the encoder prioritizes historical power, whereas the decoder combines upper-level wind speeds from both NWP models, consistent with averaging out model-specific errors. These findings support a shift from single-site to scalable multi-site forecasting, reducing operational complexity while improving accuracy in large portfolios.

Figures & tables

Author	Model	NWP	Horizon	Scope	Sites	Static Cov.
Mi et al. (2025)	TCN-TFT	-	1h	Local	1	-
Liu et al. (2024)	TFT	-	1h	Local	1	-
Hou and Liu (2026)	TFT	-	1h	Local	1	-
Mo et al. (2024)	Transformer	-	4 h	Local	1	-
van Heerden et al. (2023)	TFT	-	24 h	Local	1	-
Michalakopoulos et al. (2025)	TFT	-	24 h	Local	2	✓

Table 1: Literature Overview of Transformer-based Wind Power Forecasting

Past Observed	Known Future (NWP)	Static
Power	ICON-D2 wind speed at 78m	Power curve (cut-in, cut-out, rated wind speed)
	ICON-D2 wind speed at 127m	Hub height
	ICON-D2 wind speed at 184m	Rotor diameter
	HRES wind speed at 100m	Farm age at inference time
	HRES wind speed at 200m	Surface altitude above sea level

Table 2: Features used for model training

Turbine	Rated Power	Hub	Diameter	Cut-in	Cut-out	Rated Speed	Frequency
Vestas V112	3.45 MW	119 m	112 m	3 m/s	25 m/s	13 m/s	21
Enercon E-82 E2	2.0 MW	138 m	82 m	2 m/s	34 m/s	13 m/s	17
Enercon E-115	2.5 MW	149 m	115 m	3 m/s	34 m/s	12 m/s	16
Vestas V80	1.8 MW	78 m	80 m	4 m/s	25 m/s	15 m/s	16
Enercon E-70 E4	2.3 MW	57 m	71 m	3 m/s	34 m/s	15 m/s	15
Vestas V90	2.0 MW	95 m	90 m	4 m/s	25 m/s	13 m/s	15

Table 3: Synthetic wind power turbines occurring in this study

Figure 1: Distribution of farm age and surface altitude above sea level

Figure 2: Per-farm

\mathrm{R^{2}}

of the global model against the local baseline, for the synthetic (100 farms) and the real (11 farms) portfolios, each with and without static covariates. Points above the diagonal (green) favor the global model.

	Synthetic			Real
	100 (static)	100 (no static)	50 (static)	Real (static)	Real (no static)
Median rel. $\Delta\mathrm{RMSE}$ [%]	$+2.42$	$-3.04$	$+2.01$	$+1.64$	$-6.19$
$p$ -value	$2.6\!\times\!10^{-10}$	$1.8\!\times\!10^{-7}$	$8.9\!\times\!10^{-6}$	$0.039$	$0.0074$
Significance	∗∗∗	∗∗∗	∗∗∗	∗	∗∗

Table 4: Two-sided Wilcoxon signed-rank test on

\mathrm{RMSE}

against the local baseline. Green/red columns denote a statistically significant improvement/degradation of the global model relative to the local baseline (

p<0.05

); all five comparisons remain significant after Holm–Bonferroni correction. (

{}^{*}p<0.05

{}^{**}p<0.01

{}^{***}p<0.001

Figure 3: Feature importance weights of the VSN for the model trained on synthetic (left) and real (right) data.

	Test Period 1 (100 farms)			Test Period 2 (100 farms)			Test Period 1 (50 farms)			Test Period 2 (50 farms)
	Aug–Oct 2025			Nov 2025–Jan 2026			Aug–Oct 2025			Nov 2025–Jan 2026
Model	$\mathrm{R^{2}}$	$\mathrm{RMSE}$	$\mathrm{MAE}$	$\mathrm{R^{2}}$	$\mathrm{RMSE}$	$\mathrm{MAE}$	$\mathrm{R^{2}}$	$\mathrm{RMSE}$	$\mathrm{MAE}$	$\mathrm{R^{2}}$	$\mathrm{RMSE}$	$\mathrm{MAE}$
Global-100 (static)	86.2	8.0	4.7	79.6	17.7	10.7	85.8	8.3	4.9	79.9	18.0	11.1
Global-100 (no static)	84.2	8.4	5.0	76.6	18.6	11.3	84.1	8.7	5.2	77.2	19.0	11.7
Global-50 (static)	–	–	–	–	–	–	85.5	8.3	5.1	79.8	18.1	11.2
Local	85.2	8.3	5.1	78.7	18.0	11.3	84.7	8.6	5.3	79.1	18.3	11.5

Table 5: Normalized performance on the synthetic portfolio across both test periods, for the 100- and 50-farm evaluation sets. Best

\mathrm{R^{2}}

per column in bold; “–” marks configurations not evaluated on the respective set.

	Test Period 1			Test Period 2
	Dec 2023–Feb 2024			Mar–May 2024
Model	$\mathrm{R^{2}}$	$\mathrm{RMSE}$	$\mathrm{MAE}$	$\mathrm{R^{2}}$	$\mathrm{RMSE}$	$\mathrm{MAE}$
Global (static)	76.2	15.3	10.7	72.7	11.6	7.7
Global (no static)	72.6	16.6	12.2	68.1	12.5	8.3
Local	74.2	16.2	11.5	72.1	11.8	8.0
LinReg	63.5	19.2	14.3	62.2	13.6	10.3

Table 6: Normalized performance on the real portfolio (11 farms) across both test periods. Best

\mathrm{R^{2}}

per period in bold.

Hyperparameter	Value
batch_size	46
learning_rate	$1.95\times 10^{-4}$
n_heads	8
hidden_dim	40
dropout	0.0992
num_lstm_layers	1

Table 7: Optimal hyperparameters of the 100-farm model.

Data	Model	Duration ${}_{\mathrm{TP}_{1}}$	Duration ${}_{\mathrm{TP}_{2}}$	$N_{\mathrm{TP}_{1}}$	$N_{\mathrm{TP}_{2}}$
Synthetic	100 static	2 h 12 m	1 h 46 m	292k	329k
	100 no static	2 h 12 m	1 h 46 m	292k	329k
	50 static	55 m	1 h 13 m	146k	164k
	Local ( $\times$ 100)	42 m	42 m	$\approx$ 2.9k	$\approx$ 3.3k
Real	Global (static)	1 m 23 s	4 m 13 s	5.3k	9.3k
	Global (no static)	1 m 16 s	3 m 47 s	5.3k	9.3k

Table 8: Training cost per configuration on a single RTX A6000, split by test period (

\mathrm{TP}_{1}

\mathrm{TP}_{2}

). Each sample comprises a known-input window of

96\times 5

, an observed window of

48\times 1

, 7 static features, and a 48-step target. For the local models,

N

is the per-site training-set size (the data is partitioned across the sites), whereas the listed duration is the total wall-clock for all sites trained concurrently on one GPU (4 parallel runs at 0.25 GPU each for the synthetic and 10 at 0.1 GPU each for the real data); all other configurations use a full GPU.

Wind Power Forecasting

Wind power forecasting predicts turbine and farm output from operational data, weather forecasts, and remote observations to support reliable energy planning.

See all 27 papers

Jun 7, 2026

Tyan-WP: A Wind Power Foundation Model for Ultra-Short-Term Probabilistic Forecasting

Global wind power capacity, especially in China, is booming, with new farms spanning diverse terrains and climates. The industry urgently needs accurate wind power foundation models to shorten commissioning and accelerate grid connection. This is because site-specific time series models (TSMs) are not well suited to data-scarce scenarios and generalize poorly, while generic large time series models (LTSMs) are mostly limited to univariate inputs and cannot fully exploit static site attributes or the dependencies between power and meteorological covariates, leading to insufficient accuracy. To fill this gap, we propose Tyan-WP, the first wind power foundation model for ultra-short-term probabilistic forecasting. Pretrained on a large-scale wind power dataset covering more than 126,000 U.S. sites over seven years, Tyan-WP further improves zero-shot forecasting through two domain-specific module designs: static site embedding using coordinate, terrain, and ecoregion metadata, and a power-aware meteorological fusion (PAMF) module that models interactions between historical power and meteorological covariates. Under a unified evaluation protocol, Tyan-WP surpasses eight site-specific supervised TSMs on 10 in-domain sites and outperforms eleven generic LTSMs on 127 in-domain sites, reducing MAE by 19.9%, RMSE by 16.6%, CRPS by 22.2%, and AQL by 21.7%, while raising R^2 by 16.7%. It further demonstrates strong cross-geography generalization on six real U.K. sites. These results show that the wind power foundation model can achieve accurate zero-shot forecasting without target-site training, providing a practical pathway for rapid turbine onboarding and probabilistic risk management at new wind farms.

Jiahui Huang, Ao Luo, Lei Liu +6

School of Information Science and Technology, University of Science and Technology of China, Hefei, 230022, China · China Electric Power Research Institute, Beijing, 100192, China

Zero-Shot Time Series Forecasting Time Series Forecasting Wind Power Forecasting Time-Series Foundation Models Probabilistic Time Series Forecasting Power Systems

Ask PDF

Jul 19, 2026

Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction

Accurate short-term wind power forecasting is essential for grid stability and operational planning, yet remains challenging due to the complex interactions between atmospheric conditions and turbine dynamics. However, existing methods fail to effectively incorporate weather forecasting with wind turbine data (i.e., SCADA), leading to suboptimal solutions. To address this, we introduce a multimodal framework that integrates historical point-based SCADA data with grid-based Numerical Weather Prediction (NWP) forecasts, which is challenging due to heterogeneous input and the complex physical wind-turbine interactions. Our approach first explicitly decomposes inputs into scalar and vector features to better capture both site-specific and geometric dependencies and then incorporates a geometric encoder to extract rotation-invariant features from wind vectors. We further leverages a Fourier Neural Operator (FNO) architecture, which performs global convolutions in the frequency domain to efficiently model long-range spatiotemporal relationships. Extensive experiments on three real-world wind farms, with weather forecasting data, demonstrate that our model consistently outperforms state-of-the-art baselines, highlighting the effectiveness of its physically-informed design. The core implementation of our method is publicly available at: https://github.com/shawn-sypiao/GWPF.

Shiyuan Piao, Fan Zehui, Yang Liu +4

The Hong Kong University of Science and Technology · The Chinese University of Hong Kong · Tsinghua University Goldwind Science and Technology Co.,Ltd

Multivariate Time Series Forecasting Time Series Forecasting Spatiotemporal Forecasting Wind Power Forecasting

Ask PDF

Oct 18, 2026

Multi-site Learning for Forecasting Day-Ahead Wind-Power Production for Onshore Wind Farms in Great Britain with Deep Neural Networks

Accurate day-ahead wind-power production forecasting is increasingly important for energy system operations, market activity and balancing Great Britain's grid as wind generation grows. In this study, we train multi-site deep neural networks (DNNs) using pooled historical production and reanalysis weather data from 88 onshore wind Balancing Mechanism Units (BMUs). Forecasts are produced at 09:00 UK time for the following midnight-to-midnight delivery day, matching NESO's operational day-ahead forecasting schedule. We use reanalysis weather as future covariates during training and numerical weather prediction (NWP) forecasts during inference, extending training beyond the available historical NWP archive. Our multi-site recurrent neural network outperforms site-specific RNNs and a multi-site XGBoost model, while processing the full future-weather trajectory for the forecast horizon further improves accuracy relative to hour-aligned approaches. Our best model, an RNN with a bidirectional weather encoder (BiRNN) for NWP trajectory processing, reduces MW mean absolute error and root mean squared error relative to NESO by 17.94% and 12.51%, respectively, across 638,246 matched BMU-hour observations under conditions approximating unconstrained wind farm operations. In 40 leave-one-out trials, the BiRNN achieves lower MAE than NESO at 32 excluded sites, indicating transferable forecasting ability despite reduced accuracy when sites are omitted from training. Our capacity-estimation, data-cleaning and reanalysis-based training approach provides a basis for scaling NWP-aware multi-site forecasting to additional wind farms, longer production histories, and increased forecast accuracy beyond that seen in this study.

James Kean, Aidan O'Sullivan

UCL Energy Institute, University College London

Wind Power Forecasting Time Series Forecasting Time Series Forecasting with Covariates Recurrent Neural Networks

Ask PDF

Abstract

Wind Power Forecasting

Explore similar work

Tyan-WP: A Wind Power Foundation Model for Ultra-Short-Term Probabilistic Forecasting

Fourier Geometric Wind Power Forecasting with Numerical Weather Prediction

Multi-site Learning for Forecasting Day-Ahead Wind-Power Production for Onshore Wind Farms in Great Britain with Deep Neural Networks