Adaptive expert fusion model for online wind power prediction

Xiufeng Liu, Renfang Wang, Jingtong Wu, Xu Cheng, Xiufeng Liu 0001, Hong QiuNeural Networks 2025 · DOI

TL;DR

This paper introduces an "Adaptive Expert Fusion Model" (EFM+) for more accurate and real-time wind power prediction. By dynamically combining predictions from multiple specialized models and adapting their weights based on recent performance, EFM+ significantly improves accuracy and robustness, crucial for integrating variable wind energy into power grids.

What Problem Does This Paper Address?

Wind energy is vital for a sustainable future, but its inherent variability and uncertainty pose significant challenges for grid stability and efficient operation. Accurate and timely wind power prediction is essential for optimal power system planning, generation scheduling, and load balancing. Existing prediction models often struggle with the dynamic nature of wind conditions, leading to inaccurate forecasts, especially for intra-hour scheduling. Many models rely on static relationships or global performance metrics, failing to adapt to rapid changes or local data distributions. This paper addresses the need for an adaptive, real-time prediction model that can handle non-stationarity and uncertainty in wind power generation.

Methods

The core of the proposed solution is the Adaptive Expert Fusion Model (EFM+). It's an ensemble model that combines predictions from multiple "expert" models. The key innovations are:

  • XGBoost and Self-Attention LSTM as Experts: The model leverages XGBoost for its ability to capture complex, non-linear relationships and self-attention LSTM for its strength in modeling temporal dependencies in time-series data.
  • Dynamic Weighting Scheme: Instead of fixed weights, EFM+ uses a novel dynamic weighting scheme. This scheme assigns weights to expert models based on their performance on recent, similar data samples. This allows the model to adapt to concept drift and local data variations.
  • Subset-Based Error Calculation: To determine these dynamic weights, the model calculates errors not on the entire dataset, but on a subset of similar samples found using a k-nearest neighbor approach. This provides a more localized and relevant performance assessment.
  • Online Learning and Recurrent Batch Normalization: The model is designed for online prediction, meaning it continuously updates its parameters as new data arrives. Recurrent batch normalization is used to stabilize training.
  • Adaptive Forecasting Component: This component, featuring an "expert gate" and a "mixer gate," intelligently selects and combines the outputs of the individual models and weighted features.

Datasets

The authors used a real-world wind farm dataset provided by the Baidu KDD Cup. This dataset includes spatial and temporal information from 134 wind turbines, with data collected at 10-minute intervals. It contains parameters like wind speed, pitch angle, rotational speed, and active power output, spanning one year (January 1, 2022, to December 31, 2022).

Key Insights and Ideas

  • Local Performance Matters: The most significant insight is that adapting model weights based on performance on recent, similar data is far more effective than relying on global performance metrics or static weights. This allows the model to be highly responsive to changing wind conditions.
  • Ensemble Diversity is Key, But Not Too Much: The study found that combining multiple diverse expert models is beneficial, but there's a sweet spot. Too few models limit the ensemble's adaptability, while too many can introduce noise and degrade performance.
  • Clustering Improves Adaptability: Preprocessing the data using K-means clustering significantly improved the performance of all models, including EFM+, by reducing data interference and helping the model capture complex wind patterns more effectively.
  • Real-time Adaptability is Crucial for Intra-Hour Scheduling: The paper strongly emphasizes that for effective intra-hour scheduling, models must not only be accurate but also adapt in real-time to sudden changes in wind power generation.

Key Results

  • Superior Accuracy: EFM+ consistently outperformed all baseline methods, including other state-of-the-art models, across various metrics (RMSE, MAE, MAPE). For example, EFM+ achieved an RMSE of 33.37 and MAE of 28.31, significantly lower than other methods.
  • Robustness and Stability: The model demonstrated high robustness and stability across different scenarios, including varying noise levels and feature sets.
  • Effective Feature Selection: The study identified wind speed, pitch angle, and reactive power as the most influential features for wind power prediction.
  • Computational Cost: While highly accurate, EFM+ has a higher computational cost compared to simpler models, which is a trade-off for its advanced adaptive capabilities.

Limitations and Future Work

  • Computational Complexity: The authors acknowledge that the online learning aspect of EFM+ can lead to increased resource consumption, especially with large datasets, posing a challenge for deployment on resource-constrained devices.
  • Similarity Measure: The choice of Euclidean distance as the similarity measure for finding similar samples might not be optimal in all cases.
  • Hyperparameter Tuning: Model performance is still dependent on hyperparameter selection, requiring fine-tuning for different datasets.

For future work, the authors plan to:

  • Extend EFM+ to predict other renewable energy sources (solar, hydro) and integrate them into a unified framework.
  • Incorporate additional data sources like weather forecasts, grid load, and market prices.
  • Develop more efficient and scalable algorithms for online learning and adaptive forecasting.
  • Explore more interpretable methods for model evaluation.

Summary updated

BibTeX
@article{liu2025adaptive,
  author = {Xiufeng Liu and Renfang Wang and Jingtong Wu and Xu Cheng and Xiufeng Liu and Hong Qiu},
  title = {{Adaptive expert fusion model for online wind power prediction}},
  journal = {Neural Networks},
  year = {2025},
  doi = {10.1016/j.neunet.2024.107022},
}
Figure