Hybrid Semantic Context-Enhanced Ensemble Learning for Wind Power Ramp-Event Forecasting and Uncertainty-Aware Evaluation

Momina Liaqat Ali; Muhammad Abid; Muhammad Abdullah; Aneela Zameer

cs.LGAug 29, 2026

Hybrid Semantic Context-Enhanced Ensemble Learning for Wind Power Ramp-Event Forecasting and Uncertainty-Aware Evaluation

Authors: Momina Liaqat Ali, Muhammad Abid, Muhammad Abdullah, Aneela Zameer

Organizations: Department of Industrial and Systems Engineering, University of Tennessee, Knoxville Knoxville, TN 37996, USA. · Department of Mechanical and Aerospace Engineering, University of Tennessee, Knoxville Knoxville, TN 37996, USA. · Department of Mechanical Engineering, Pakistan Institute of Engineering and Applied Sciences, Islamabad, Pakistan. · Department of Computer Science, Pakistan Institute of Engineering and Applied Sciences, Islamabad, Pakistan.

Abstract

Wind power ramp events which are sudden, large swings in turbine output over short windows are difficult to estimate, and standard models often miss them. Hybrid forecasting approach is built which augments semantic context to ramp-event forecast. Rather than applying an extensive language model directly to predict turbine operating data, we have implemented a pipeline where turbine operating data is converted to simplified text, which is then converted to dense embeddings to be used as inputs for ensemble models incorporated with other features. Testing runs are performed at multiple intervals within the SDWPF dataset, including 10-minute, 30-minute, and 60- minute horizons, with ramp events constituting the highest change in future power output. We check robustness against autoregressive, LSTM, and GRU baselines plus several ensemble configurations, using Diebold-Mariano tests and bootstrap confidence intervals, and we vary the ramp threshold, compress the embeddings with PCA, and validate externally on Kaggle SCADA and NREL data with uncertainty-aware scoring. The semantic-context features produce negligible yet statistically significant gains over the baselines in multiple paired ensemble runs, most clearly at the 30- and 60-minute horizons where these gains hold across different ramp-threshold definitions, and PCA compression helps in some longer-horizon cases. The best context- augmented ensembles rank near the top overall, though the GRU model still posts the lowest ramp-event RMSE at 30 and 60 minutes. External tests confirm the error reduction generalizes across datasets, but the size of the gain depends on both model and dataset. Prediction intervals cover most test cases well but weaken during ramp events, pointing to a localized shift in the data distribution.

Figures & tables

Figure 1: Detailed workflow of the suggested combined semantic-contextual based wind power estimation system. The design covers data preparation, construction of numerical/semantic contextual content, modeling enhancements, verification of statistical measures, ramping sensitivity study, external proof of the model, and calculation of uncertainties.

Research stream	Recent focus	Position of the present study
Wind power forecasting	Statistical, machine-learning, deep-learning, hybrid, and probabilistic forecasting	Evaluates classical, ensemble, sequence-learning, and context-enhanced models across multiple horizons
Ramp-event forecasting	Ramp detection, regression, classification, probabilistic prediction, attention, graph-based models, and multi-task learning	Defines horizon-specific ramp subsets and evaluates forecasting performance separately during high-ramp conditions
Feature representation	Historical measurements, meteorological inputs, lagged features, statistical features, feature extraction, and learned representations	Adds symbolic operating states, TF-IDF features, LLM-derived semantic embeddings, and PCA-compressed semantic representations
Transformer forecasting	Attention, decomposition, frequency-domain modeling, patch-based representations, exogenous variables, and multiscale modeling	Uses strong sequence-learning models as comparison baselines for context-enhanced forecasting
LLM and foundation models	LLM reprogramming, autoregressive LLM forecasting, zero-shot forecasting, and pretrained time-series foundation models	Uses an LLM as a semantic operating-context encoder rather than as the direct forecasting model
Uncertainty estimation	Probabilistic forecasting, quantile prediction, conformal intervals, and distributional forecasting	Evaluates interval coverage and width separately for complete-test and ramp-event samples

Table 1: Positioning of the present study relative to recent literature.

Item	Description
Primary dataset	Spatial Dynamic Wind Power Forecasting (SDWPF) dataset
Original turbine coverage	134 wind turbines
Turbines used in main study	Turbines 1–20
External validation datasets	Kaggle wind turbine SCADA and NREL wind turbine data
Sampling interval	10 minutes
Forecast horizons	10, 30, and 60 minutes

Table 2: Summary of dataset and forecasting setup.

Figure 2: Spatial layout of the SDWPF wind turbines. Filled markers indicate turbines 1–20 used in the main experiments, while open markers indicate the remaining turbines in the wind farm.

Variable group	Variables	Forecasting role
Meteorological variables	Wind speed, wind direction, external temperature	Capture atmospheric conditions
Turbine operating variables	Internal temperature, nacelle direction, blade pitch angles	Represent turbine behavior
Power variables	Active power, reactive power	Represent current generation and electrical state
Spatial and temporal variables	Turbine coordinates, hour, minute, day/month proxies	Capture location and time patterns
Lagged and rolling variables	Power lags, rolling means, rolling standard deviations	Capture recent dependence and volatility
Context variables	Wind-speed change, power change, direction change, variability indicators	Capture operating-regime changes

Table 3: Main input variable groups used for wind power forecasting.

Notation	Description
$i$	Turbine index
$t$	Current time step
$h$	Forecast horizon, where $h\in\{1,3,6\}$
$P_{i,t}$	Active power output of turbine $i$ at time $t$
$\hat{P}_{i,t+h}$	Predicted active power output at horizon $h$
$y_{i,t}^{(h)}$	Forecasting target, defined as $P_{i,t+h}$

Table 4: Mathematical notation used in the forecasting formulation.

Forecast horizon	Prediction step	Ramp definition	Ramp threshold
10 min	$h=1$	Top 10% of $\|P_{i,t+1}-P_{i,t}\|$	93.552 kW
30 min	$h=3$	Top 10% of $\|P_{i,t+3}-P_{i,t}\|$	172.700 kW
60 min	$h=6$	Top 10% of $\|P_{i,t+6}-P_{i,t}\|$	248.814 kW

Table 5: Main ramp-event definition and horizon-specific thresholds.

Context representation	Construction	Forecasting role
Numerical context	Wind-speed change, power change, direction change, and variability features	Captures short-term operating changes
Symbolic context	Discretized wind, power, change, and ramp-risk categories	Represents operating regimes in categorical form
TF-IDF text context	Short descriptions created from symbolic states	Tests sparse textual context representation
Rich semantic context	Embeddings of detailed operational summaries	Captures semantic turbine-state and ramp-risk information
PCA-compressed context	Lower-dimensional projection of rich semantic embeddings	Tests whether compressed semantic features reduce embedding noise

Table 6: Context representations constructed in this study.

	$\displaystyle\mathbf{x}_{i,t}^{\mathrm{NoLLM}}$	$\displaystyle=\text{numerical and contextual features},$
	$\displaystyle\mathbf{x}_{i,t}^{\mathrm{RichLLM}}$	$\displaystyle=\left[\mathbf{x}_{i,t}^{\mathrm{NoLLM}},\mathbf{z}_{i,t}^{\mathrm{LLM}}\right],$
	$\displaystyle\mathbf{x}_{i,t}^{\mathrm{PCA}(d)}$	$\displaystyle=\left[\mathbf{x}_{i,t}^{\mathrm{NoLLM}},\mathbf{z}_{i,t}^{\mathrm{PCA}(d)}\right].$

Algorithm 1 Hybrid Semantic-Context-Enhanced Ramp Forecasting Framework

Evaluation setting	Samples included	Purpose
All-test evaluation	All observations in the chronological test period	Measures general forecasting performance
Ramp-event evaluation	Top 10% highest absolute future power-change samples	Measures performance under high-variability regimes
Ramp-threshold sensitivity	Top 10%, 15%, and 20% highest absolute power-change samples	Tests robustness to ramp definition
Persistence comparison	Model RMSE compared with persistence RMSE	Quantifies improvement over a strong short-horizon baseline
Paired LLM comparison	Same model trained with and without rich semantic embeddings	Isolates the contribution of LLM-derived context
Bootstrap confidence intervals	Resampled paired test predictions	Quantifies uncertainty in RMSE gains

Table 7: Evaluation settings used in the experiments.

Model category	Models included	Purpose
Persistence	Persistence	Strong short-horizon reference baseline
Classical time-series baselines	Rolling mean, AR-Ridge, AR-ElasticNet	Tests autoregressive lag-based forecasting performance
Context-ablation XGBoost	XGBoost-Lag, Symbolic-XGBoost, TF-IDF-XGBoost, Short LLM-XGBoost, Rich LLM-XGBoost	Tests different context representations
Expanded ML baselines	Ridge, ElasticNet, Random Forest, Extra Trees, HistGradientBoosting, XGBoost, LightGBM, CatBoost	Compares against strong non-LLM models
Paired LLM comparison	Same model trained with and without rich semantic embeddings	Isolates the effect of LLM-derived context
PCA embedding variants	Rich-LLM-384, PCA-16, PCA-32, PCA-64, PCA-128	Tests whether compressed semantic embeddings reduce noise

Table 8: Summary of model categories used in the experiments.

Figure 3: Ramp-event RMSE across context-ablation forecasting models at 10-, 30-, and 60-minute forecast horizons.

Figure 4: Ramp-event RMSE improvement relative to persistence for context-ablation models. Positive values indicate lower RMSE than persistence.

Horizon	Model	MAE	RMSE	$R^{2}$	RMSE gain (%)
10 min	Persistence	167.17	187.32	0.7439	0.00
10 min	XGBoost-Lag	156.48	182.05	0.7581	2.82
10 min	Symbolic Context-XGBoost	154.71	180.53	0.7621	3.63
10 min	TF-IDF TextContext-XGBoost	154.83	180.83	0.7613	3.47
10 min	Short LLMContext-XGBoost	154.88	180.76	0.7615	3.51
10 min	Rich LLMContext-XGBoost	153.77	179.03	0.7660	4.43

Table 9: Ramp-event forecasting ablation across context representations.

Horizon	Subset	Context RMSE	Rich LLM RMSE	Gain (%)	DM stat.	$p$ -value
10 min	All-test	67.022	65.850	1.749	15.919	$<0.001$
10 min	Ramp-event	180.529	179.027	0.832	6.170	$<0.001$
30 min	All-test	121.869	117.211	3.822	23.477	$<0.001$
30 min	Ramp-event	290.656	288.570	0.718	5.067	$<0.001$
60 min	All-test	167.212	166.700	0.306	1.905	0.057
60 min	Ramp-event	383.094	383.396	-0.079	-0.545	0.586

Table 10: Diebold-Mariano test results for Context-XGBoost versus Rich LLMContext-XGBoost.

Horizon	Model	MAE	RMSE	$R^{2}$	RMSE gain (%)
10 min	XGBoost Standard	152.731	178.848	0.767	4.524
10 min	Extra Trees	155.468	179.691	0.764	4.074
10 min	CatBoost	154.824	179.719	0.764	4.059
10 min	LightGBM	154.101	180.366	0.763	3.714
10 min	HistGB	154.435	180.433	0.762	3.678
30 min	CatBoost	249.109	287.395	0.423	8.499

Table 11: Top five ramp-event models in the expanded model comparison.

Horizon	Model	No-LLM RMSE	Rich-LLM RMSE	LLM gain (%)
10 min	Random Forest	179.775	178.753	0.569
10 min	CatBoost	179.564	178.840	0.403
10 min	XGBoost	179.233	178.971	0.146
10 min	Extra Trees	179.339	179.463	-0.069
10 min	HistGB	180.411	180.550	-0.077
30 min	HistGB	293.396	289.110	1.461

Table 12: Effect of adding rich LLM-derived embeddings to different model families on SDWPF ramp-event RMSE.

Figure 5: Bootstrap 95% confidence intervals for Rich-LLM RMSE gain over the corresponding no-LLM model. Positive values indicate lower ramp-event RMSE after adding rich LLM-derived embeddings.

Horizon	Model	RMSE gain (%)	95% CI lower	95% CI upper
10 min	Random Forest	0.569	0.128	0.989
10 min	CatBoost	0.403	0.216	0.583
10 min	XGBoost	0.146	-0.147	0.473
30 min	HistGB	1.461	1.199	1.725
30 min	XGBoost	1.135	0.878	1.388
30 min	Random Forest	0.804	0.447	1.160

Table 13: Bootstrap confidence intervals for paired Rich-LLM ramp-event RMSE gain.

Figure 6: Sensitivity of best Rich-LLM RMSE gain to ramp-event definition. Positive values indicate lower RMSE after adding rich LLM-derived embeddings.

Horizon	Ramp definition	Best model	RMSE gain (%)
10 min	Top 10%	Random Forest	0.569
10 min	Top 15%	Random Forest	0.823
10 min	Top 20%	Random Forest	0.979
30 min	Top 10%	HistGB	1.461
30 min	Top 15%	HistGB	1.651
30 min	Top 20%	XGBoost	1.814

Table 14: Ramp-threshold sensitivity of the best Rich-LLM RMSE gain.

Horizon	Model	Best semantic variant	Best RMSE	Gain vs. No-LLM (%)
10 min	Random Forest	Rich-LLM-384	179.795	1.334
10 min	HistGB	Rich-LLM-384	180.316	0.553
10 min	XGBoost	PCA-64	180.323	0.386
10 min	CatBoost	Rich-LLM-384	179.159	0.295
30 min	Random Forest	Rich-LLM-384	285.783	1.568
30 min	XGBoost	Rich-LLM-384	288.725	1.077

Table 15: Summary of Rich-LLM and PCA-compressed semantic embedding performance.

Horizon	Model group	MAE	RMSE	sMAPE	$R^{2}$
10 min	Rich-LLM ensemble	154.555	178.753	30.730	0.767
10 min	No-LLM ensemble	152.766	179.233	31.060	0.766
10 min	GRU	158.040	182.352	32.448	0.757
10 min	LSTM	154.636	182.350	30.854	0.757
10 min	Persistence	167.170	187.323	–	0.744
10 min	AR-Ridge	–	202.601	–	–

Table 16: Final ramp-event model comparison including classical, ensemble, and deep sequence baselines.

Dataset	Horizon	Non-LLM RMSE	LLM RMSE	Best LLM model	LLM gain (%)
Kaggle SCADA	10 min	262.889	262.502	HistGB	0.147
Kaggle SCADA	30 min	419.826	417.090	Random Forest	0.652
Kaggle SCADA	60 min	564.367	554.269	Extra Trees	1.789
NREL Wind	10 min	10.354	10.046	Extra Trees	2.967
NREL Wind	30 min	15.084	13.483	Extra Trees	10.615
NREL Wind	60 min	20.755	18.305	Extra Trees	11.805

Table 17: External validation of best non-LLM and best LLM-augmented models. Positive gain indicates lower overall RMSE for the LLM-augmented model.

Figure 7: External validation results showing overall RMSE gain from adding rich LLM-derived operational context on Kaggle SCADA and NREL wind datasets. Positive values indicate lower RMSE for the LLM-augmented model compared with the best non-LLM counterpart.

Figure 8: Feature-importance contribution of rich LLM-context embeddings across forecast horizons.

Figure 9: Prediction interval coverage probability for all-test and ramp-event samples. The nominal prediction interval is 80%.

Figure 10: Mean prediction interval width for all-test and ramp-event samples.

Horizon	All PICP	Ramp PICP	Cov. drop	All MPIW	Ramp MPIW	Width inc.
10 min	87.74	46.80	40.94	120.88	293.41	142.74%
30 min	87.07	45.89	41.18	228.75	478.98	109.39%
60 min	85.19	47.12	38.07	321.44	623.88	94.09%

Table 18: Uncertainty evaluation for Rich LLMContext quantile XGBoost.

Explore similar work

Feb 13, 2026cs.LG

Probabilistic Wind Power Forecasting with Tree-Based Machine Learning and Weather Ensembles

Accurate production forecasts are essential for the integration of renewable energy sources into the power grid. This paper illustrates how to obtain probabilistic forecasts of wind power generation using gradient boosting trees and an ensemble of weather forecasts. To this end, we perform a comparative analysis across three state-of-the-art probabilistic prediction methods-conformalized quantile regression, natural gradient boosting and conditional diffusion models-all of which can be combined with tree-based machine learning. The methods are validated using four years of data for all Belgian offshore wind farms. We benchmark the models against the power curve and a calibrated wake model as well as a probabilistic method using stochastic variational Gaussian process regression. The tree-based models significantly reduce the mean absolute error in comparison to the deterministic baselines. Additionally, all three methods outperform the Gaussian process baseline in probabilistic skill, while two out of the three also improve point forecast accuracy. The conditional diffusion model attains the best performance, with improvements of 5% in mean absolute error and 12% in continuous rank probability score compared to the probabilistic baseline. Last, the results indicate an average improvement in point forecast accuracy of 17% by using an ensemble of weather forecasts instead of a single provider.

Max Bruninx, Diederik van Binsbergen, Timothy Verstraeten +2

Vrije Universiteit Brussel, Brussels, Belgium · Norwegian University of Science and Technology, Trondheim, Norway

Gradient-Boosted Decision Trees Ensemble Learning Time Series Forecasting Conformal Prediction Wind Power Forecasting Conditional Diffusion Models+2

Ask PDF

Aug 27, 2026cs.LG

Technical Comparative Benchmarking Study: Advanced AI Hybrid Methods for Renewable Energy Farm Optimization and Forecasting

This study provides a comprehensive benchmarking of conventional machine learning (ML), ensemble learning, deep neural networks, recurrent architectures, Transformers, graph based models, and hybrid ensemble deep learning approaches under complementary renewable energy scenarios. Three datasets are considered: a large scale WEC dataset, a 16 WEC dataset, and operational 10 min SCADA measurements at the Penmanshiel wind farm. For structured WEC layout data, tree ensembles exhibited a clear advantage over conventional ML and neural predictors because randomized partitioning and boosting efficiently captured nonlinear layout power interactions without requiring explicit feature representation learning. The Extra Trees was the strongest model, achieving considerable results. Relative to the MLP baseline, this corresponds to an approximately 63.7% reduction in MAE, demonstrating the suitability of randomized tree ensembles for high dimensional structured WEC data. Also, STGCN reduced the MAE to approximately 167.0 kW and achieved R = 0.93 by explicitly learning spatial and temporal turbine interactions. The best overall forecasting accuracy was obtained by the RF BiLSTM hybrid, with an MAE=150.5 kW. Compared with standalone LSTM, this represents an approximately 75% reduction in MAE, while improving on STGCN by approximately 10.0%. Finally, the experiments reveal that no single AI architecture is universally optimal: randomized and boosted ensembles are particularly effective for structured WEC surrogate modeling, graph networks become advantageous when explicit spatial interactions dominate, and ensemble recurrent hybrids provide the strongest balance when nonlinear tabular relationships and temporal dynamics coexist.

Majid Masoumi, Asghar Dashtiy, Mohammad Dehghan +1

Department of Computer Science Yazd University Yazd · Department of Mechanical Engineering Yazd University Yazd

Wind Power Forecasting Forecasting Benchmarks Ensemble Learning Neural Surrogate Modeling Graph Neural Networks

Ask PDF

Jun 7, 2026cs.LG

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 \textbf{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