Enhancing spatiotemporal wind power forecasting with meta-learning in data-scarce environments

Xiufeng Liu, Renfang Wang, Jingtong Wu, Xu Cheng, Xiufeng Liu 0001, Hong QiuEngineering Applications of AI 2025 · DOI

TL;DR

Accurate wind power forecasting is crucial for grid stability but challenging due to wind's variability and limited data for new farms. This paper introduces MAML-STALSTM, a novel framework that combines self-attention enhanced Spatio-Temporal LSTMs with Model-Agnostic Meta-Learning (MAML) to achieve highly accurate and adaptable wind power forecasts, even with scarce data.

What Problem Does This Paper Address?

Integrating wind power into the grid is vital for sustainability, but its inherent intermittency poses a significant challenge. Accurate wind power forecasting (WPF) is essential for grid stability and efficient operation. However, traditional forecasting methods struggle with the complex, spatiotemporal nature of wind patterns. Furthermore, deploying new wind farms or dealing with changing conditions often means limited historical data, making it difficult to train complex deep learning models effectively. Existing approaches either excel at capturing spatiotemporal dynamics but require vast amounts of data, or they are data-efficient but lack the ability to model complex spatiotemporal interactions. This paper aims to bridge this gap by developing a model that is both adept at understanding complex wind patterns and capable of rapid adaptation with minimal data.

Methods

The core of this work is the MAML-STALSTM framework, which ingeniously combines two key components:

  • Self-Attention Enhanced Spatio-Temporal LSTM (ST-LSTM): This is built upon the PredRNN architecture.
  • ST-LSTM: Extends standard LSTMs to capture spatial relationships between wind turbines in addition to temporal dependencies. This is crucial for understanding how turbines within a farm influence each other.
  • Self-Attention Mechanism: Integrated into each ST-LSTM layer, this mechanism dynamically weighs the importance of different input features and past states, allowing the model to focus on the most relevant information for forecasting. This helps in capturing long-range spatiotemporal dependencies more effectively than traditional convolutional approaches.
  • Model-Agnostic Meta-Learning (MAML): This meta-learning algorithm is used to train the model to "learn to learn."
    • Bi-Level Optimization: MAML employs an inner loop for task-specific adaptation (fine-tuning the model on a small dataset for a specific wind farm) and an outer loop for meta-parameter optimization (finding initial model parameters that are broadly adaptable to new tasks). This allows the model to quickly adapt to new wind farms or changing conditions with very few training examples.

The framework also incorporates robust data preprocessing, including cleaning, imputation, normalization, and denoising, to ensure high-quality input data.

Datasets

The authors evaluated their MAML-STALSTM framework on two distinct datasets:

  • Onshore Wind Farm Dataset: This dataset comprises hourly records from four wind turbines over five years (January 2017 to December 2021). It includes nine distinct features, such as wind speed, temperature, and wind direction.
  • Offshore Wind Farm Dataset: This dataset features high-resolution power output measurements from five wind turbines over two years (January 2022 to January 2024), with data collected at 15-minute intervals. It captures periods of regular operation and extreme wind fluctuations, making it a more challenging benchmark.

Key Insights and Ideas

  • Meta-Learning for Data Scarcity is Key: The paper strongly demonstrates that meta-learning, specifically MAML, is a powerful tool for overcoming data scarcity in wind power forecasting. It allows models to adapt rapidly to new scenarios without needing extensive historical data.
  • Spatiotemporal Awareness is Crucial: Simply capturing temporal patterns isn't enough. The ST-LSTM component highlights the importance of understanding how wind turbines interact within a farm to improve forecasting accuracy.
  • Self-Attention Enhances ST-LSTM: The integration of self-attention significantly boosts the ST-LSTM's ability to capture complex, long-range spatiotemporal dependencies, leading to better performance.
  • Offshore is Tougher, But MAML-STALSTM Holds Up: While offshore environments present greater challenges due to extreme variability, MAML-STALSTM shows remarkable adaptability and maintains strong performance, even outperforming other models in certain aspects.

Key Results

  • Superior Accuracy: MAML-STALSTM consistently outperformed nine baseline models (including MLP, GRU, BiLSTM, Transformer, GNN, ConvLSTM, and ST-LSTM) on both onshore and offshore datasets across various forecasting horizons (6, 12, and 24 hours).
  • Data Efficiency: In data-scarce scenarios (e.g., 60 days of training data), MAML-STALSTM showed significantly less performance degradation compared to traditional deep learning models. For instance, it achieved a lower MAE than most baselines trained on more data.
  • Adaptability: Ablation studies showed that MAML integration significantly improved the performance of various base models, particularly ST-LSTM.
  • Offshore Performance: On the offshore dataset, MAML-STALSTM achieved MAE and RMSE values that were approximately 69% and 67% lower than iTransformer for the 24-hour forecast, demonstrating its robustness in challenging conditions.

Limitations and Future Work

  • Computational Complexity: The authors acknowledge that MAML-STALSTM has increased computational complexity due to meta-learning and self-attention, though they argue it's a worthwhile trade-off for improved performance.
  • Hyperparameter Sensitivity: The STALSTM model, in particular, was found to be sensitive to hyperparameter configurations (learning rate, number of layers).
  • Offshore High-Wind Prediction: While strong in low-wind regions, MAML-STALSTM showed poorer prediction accuracy in high-wind areas of offshore environments, with varying error patterns across turbines.
  • Future Work:
    • Exploring more advanced meta-learning techniques (e.g., Meta-SGD).
  • Incorporating uncertainty quantification for confidence bounds.
  • Integrating Numerical Weather Prediction (NWP) data.
  • Improving computational efficiency and scalability.
  • Enhancing model interpretability.
  • Further research into offshore environments, potentially by incorporating more specific environmental features and multi-scale spatial modeling.

Summary updated

BibTeX
@article{liu2025enhancing,
  author = {Xiufeng Liu and Renfang Wang and Jingtong Wu and Xu Cheng and Xiufeng Liu and Hong Qiu},
  title = {{Enhancing spatiotemporal wind power forecasting with meta-learning in data-scarce environments}},
  journal = {Engineering Applications of AI},
  year = {2025},
  doi = {10.1016/j.engappai.2025.111121},
}
Figure