DEWP: Deep Expansion Learning for Wind Power Forecasting

Wei Fan, Yanjie Fu, Shun Zheng, Jiang Bian, Yuanchun Zhou, Hui Xiong

Introduction

Wind power is a critical but notoriously volatile renewable energy source. Accurate forecasting is essential for grid stability, but it is exceptionally difficult because generated power is influenced by a "latent composition" of multiple time-evolving factors—wind speed, direction, temperature, nacelle angle, and more.

Time-series visualizations of the diverse physical variables influencing wind power generation.

The authors identify two major gaps in existing research that hinder accuracy: first, many models focus heavily on time-series patterns while neglecting the complex, non-linear relationships between the input variables themselves. Second, standard architectures often lack the "network expressiveness" required to model the intricate dependencies found in high-resolution turbine data.

The DEWP Framework

DEWP (Deep Expansion learning for Wind Power forecasting) is a novel stack-by-stack framework that explicitly decouples and models two types of dependencies: correlations between different physical variables and temporal patterns over time.

The DEWP framework architecture, featuring variable and time expansion blocks within a residual stacking structure.

Each stack in the DEWP architecture contains three primary components:

  • Variable Expansion Block: Uses four layers of 1D-convolutional filters to capture dependencies across the variable dimension. This expands raw inputs into a hidden feature space where inter-variable relationships are more apparent.
  • Time Expansion Block: This block models temporal patterns using a backcast/forecast mechanism. Crucially, it incorporates Fourier series to constrain cyclical functions, allowing the model to better capture the inherent seasonality of wind patterns.
  • Inference Block: To map these expanded latent representations to actual power values, the authors use multi-head self-attention. This allows the model to weigh different features based on their importance to the final forecast.

The framework is tied together with doubly residue learning. Each stack "removes" what it has learned from the input (backcast residue) before passing it to the next stack, while its predictions are accumulated to form the final fine-grained forecast.

Experimental Setup

The study uses real-world data from the 'La-Haute-Borne' wind farm in France. Two separate datasets were created for Turbine R80736 and Turbine R80721. Features include wind speed, wind direction, nacelle angle, pitch angle, outdoor temperature, and hub temperature. Data was aggregated into hourly intervals, normalized, and augmented with time embeddings (month, weekday, and hour) to capture temporal context.

Key Results and Insights

DEWP was tested against a wide array of baselines, including GRU, LSTNet, TCN, Transformer, Informer, and N-BEATS. It consistently achieved the lowest Mean Absolute Percentage Error (MAPE) and Mean Squared Prediction Error (MSPE). For example, on the R80736 dataset, it improved MAPE by approximately 13.4% over the second-best performing model.

Several key design choices drive this performance:

  • Decoupling is Key: By separating variable dependency learning from temporal dependency learning, the model can optimize for each independently.
  • Seasonality via Fourier: Wind is seasonal. Using Fourier series as a basis for temporal expansion acts as a powerful inductive bias, helping the model ignore noise and focus on recurring physical patterns.
  • Attention in the Latent Space: Instead of applying self-attention directly to raw inputs, DEWP applies it to the expanded features, making it more effective at identifying driving factors.

Conclusion

While DEWP achieves state-of-the-art accuracy, the authors highlight areas for future improvement, specifically regarding computational complexity and the need for online training capabilities. The use of multi-head self-attention introduces $O(N^2)$ complexity, which could become a bottleneck for massive wind farm arrays. Future directions include developing real-time update algorithms and integrating more efficient attention mechanisms to reduce the computational footprint.

Summary updated

BibTeX
@article{fan2024dewp,
  author = {Wei Fan and Yanjie Fu and Shun Zheng and Jiang Bian and Yuanchun Zhou and Hui Xiong},
  title = {{DEWP: Deep Expansion Learning for Wind Power Forecasting}},
  journal = {ACM Trans. Knowl. Discov. Data},
  year = {2024},
  doi = {10.1145/3637552},
}
Figure