A Systematic Evaluation of Current Architectures in Wind Power Forecasting

Abstract: Interval wind speed forecasting is essential for the efficient integration of wind energy into power systems, as it accounts for the inherent uncertainty of wind resources. This study presents a systematic literature review focused on hybrid approaches to interval forecasting of wind generation, exploring the combination of deep learning, modal decomposition, and statistical methods. To guide the paper selection, Latent Dirichlet Allocation (LDA) was applied for topic modeling, enabling the identification of patterns and research trends. The findings emphasize that integrating hybrid models with decomposition techniques-such as Variational Mode Decomposition (VMD) and Ensemble Empirical Mode Decomposition (EEMD)-enhances forecast accuracy and reliability by narrowing prediction intervals without compromising coverage. Regarding interval construction, most studies adopt a dual-model strategy, independently forecasting the lower and upper bounds. Input data are commonly decomposed using techniques like EMD, EEMD, or VMD, which extract frequency-based components. These components serve as inputs to models such as LSTM or ELM, trained separately for each bound. This approach allows for targeted modeling of uncertainty, improving flexibility and precision, Interval quality is typically evaluated through metrics that balance coverage and interval width. The review also highlights challenges, including the lack of standardized evaluation metrics, computational complexity, and limited real-world validation. Overall, the study reinforces the value of interval forecasting for wind energy operations and offers insights for advancing model robustness and decision-making.

• The paper demonstrates that hybrid models, combining signal decomposition methods with deep learning, significantly enhance forecasting reliability with high PICP and reduced PINAW.
• The study employs LDA-based topic discovery to map methodological trends and reveals a notable gap in the adoption of Transformer-based models for interval forecasting.
• Empirical insights highlight the benefits of multi-objective optimization in balancing forecast sharpness and coverage, while also addressing challenges in computational cost and metric standardization.

Systematic Evaluation of Wind Power Forecasting Architectures

Introduction

Accurate interval forecasting of wind power is paramount for grid integration, operational resilience, and risk-aware scheduling in modern energy systems. The reviewed work systematically maps state-of-the-art architectures for wind power interval forecasting, with explicit focus on hybrid frameworks that integrate signal decomposition, deep learning, and advanced statistical modeling. The study employs Latent Dirichlet Allocation (LDA) for topic discovery across a large corpus (2015-2024), distilling methodological trends and gaps in the literature. The review critically examines approaches for constructing, calibrating, and evaluating forecast intervals, elucidating best practices and highlighting limitations impeding broader adoption and operationalization.

Methodological Framework

The literature mapping precisely defines research questions, applies structured search and selection across English-language Scopus-indexed studies, and leverages advanced text mining and topic modeling (LDA). The resulting taxonomy distinguishes four principal forecasting paradigms: deterministic point forecasting, probabilistic/interval forecasting, deterministic interval forecasting, and hybrid deterministic approaches. Particular scrutiny is given to interval and hybrid architectures, which together comprise the core of high-impact research in wind energy forecasting.

The construction and assessment of prediction intervals are analyzed in detail, delineating explicit algorithmic strategies for the lower and upper bounds, including both symmetric (parametric) and asymmetric (nonparametric, quantile-based) formulations. Model evaluation is discussed on both point and interval performance metrics, with emphasis on those explicitly quantifying calibration and sharpness under uncertainty (PICP, CWC, PINAW, CRPS).

Key Results and Methodological Trends

Signal Decomposition and Hybridization

VMD is the most prevalent data preprocessing technique, followed by CEEMDAN and EMD. These decomposition strategies are robust for managing the nonstationarity and multi-scale characteristics of wind speed series. Hybrid approaches, integrating these decompositional frameworks with DNNs (e.g., LSTM, GRU, CNN-BiLSTM) or with kernel-based statistical learners (e.g., KDE, QR), dominate the landscape, reflecting their empirical superiority in both point and interval settings.

Multi-stage pipelines are the norm: time series are decomposed, sub-components are individually modeled (often with separate DNNs or optimized ELMs for each bound), and forecast aggregation is typically performed via weighted combination, with hyperparameter optimization schemes (e.g., particle swarm, genetic algorithms) frequently used for performance tuning.

Deep Learning and Advanced Statistical Models

LSTM networks and variants (GRU, BiLSTM, CNN-LSTM/BiLSTM) are the backbone predictors in nonlinear, multi-frequency wind data contexts. Their integration with VMD and CEEMDAN, for example, yields tangible improvements in interval sharpness and reliability compared to both stand-alone statistical time series models (ARIMA, classic QR) and traditional neural networks. Extreme Learning Machines and model ensembles further enhance robustness and interval informativeness, particularly when combined with modal decomposition and evolutionary optimization routines.

A notable gap is the non-adoption of Transformer-based models for interval forecasting in wind energy, despite emerging evidence in adjacent energy domains (see Geneva & Zabaras [52], Yang et al. [53], Hussan et al. [54]) that attention mechanisms substantially improve long-term dependency capture, scalability, and uncertainty quantification.

Interval Construction and Metrics

Interval construction methodologies fall into several archetypes:

• Symmetric, distribution-based intervals: Constructed via normality assumptions on residuals with standard deviation-based expansion (frequent in early-stage and some hybrid models).
• Asymmetric, quantile-based intervals: Estimated directly using QR, KDE, CKDE, or MCMC-derived distributions, circumventing parametric assumptions and yielding flexible, data-driven bounds.
• Direct DNN interval prediction: Multi-headed LSTM/GRU or BiLSTM networks output bounds with pinball loss or CWC-based joint optimization.

Performance measures are increasingly oriented toward interval quality: PICP is dominant for coverage, PINAW and PINRW for width normalization and sharpness, CWC as a penalized aggregate metric, and (less common) CRPS for full distributional fidelity. However, the lack of standardized benchmarking and consistent metric usage hinders direct cross-study comparison, an issue compounded by persistent reliance on point measures (RMSE, MAE) in many advanced model assessments.

Empirical Insights

Numerical results from surveyed studies demonstrate that hybrid models, especially those coupling VMD/CEEMDAN decomposition with LSTM or GRU, consistently report high PICP (>96%) and reduced width (PINAW ≤ 69%) compared to benchmarks. Multi-objective optimization further enhances the informativeness–reliability trade-off in practical settings. Asymmetric interval construction and kernel/quantile-based statistics provide enhanced flexibility and more realistic uncertainty representation in operational forecasting scenarios.

Short-term forecasting (up to 96% of studies) is the overwhelming focus, reflecting operational dispatch needs, while medium- and long-term forecasts remain underexplored due to compounded uncertainty and computational barriers.

Theoretical and Practical Implications

Hybrid interval models effectively address the nonlinearity, intermittency, and multi-scale stochasticity inherent in wind generation. They enable operators to integrate uncertainty-aware forecasts into dispatch, storage scheduling, and market bidding strategies—crucial for system resilience amidst increasing renewable penetration. The robust interval quantification supports more nuanced risk management, curtailment minimization, and improved balancing services.

The incorporation of adaptive, multi-objective optimization into interval width/coverage trade-offs makes these models operationally relevant, especially when coupled with efficient decomposition and computational strategies.

However, deployment in production systems is still stalled by several factors:

• Computational cost associated with complex hybrid and deep models.
• Nonstandardized validation and metric reporting
• Scarce real-world, real-time validation, particularly with high-frequency and heterogeneous (IoT/satellite) datasets.

Future Trajectories in AI for Wind Power Forecasting

Several research trajectories are suggested:

• Integration of Transformer-based architectures and self-attention mechanisms into interval forecasting pipelines, potentially leading to further gains in scalability, accuracy, and interval calibration, particularly at longer horizons and across spatially distributed farms.
• Expansion toward medium- and long-term interval forecasting, leveraging seasonality and multi-modal data, to meet strategic planning requirements.
• Standardization of evaluation protocols, interval forecast metrics, and cross-dataset benchmarking, facilitating transparent model comparison and operational adoption.
• Validation with real-time, multi-source data (real-world SCADA, IoT-augmented meteorological streams), ensuring model robustness and transferability.

Conclusion

The reviewed work offers a systemic, granular overview of current architectures and protocols in wind power interval forecasting, emphasizing the dominance and empirical success of hybrid decompositional–deep learning frameworks. The field is progressing toward more sophisticated uncertainty quantification, supported by advanced metrics and optimization-driven architectures. Gaps in benchmarking, interpretability, and computational scalability constrain operationalization, but the rapid incorporation of new AI paradigms and standardization efforts will likely address these deficits. Hybrid interval forecasting models will be central to the resilient and economically efficient integration of wind generation into future power systems.

Reference:

"A Systematic Evaluation of Current Architectures in Wind Power Forecasting" (2606.02849)