Short-Term Wind Power Forecasting Using Empirical Mode Decomposition and RBFNN

Zengwei Zheng, Yuanyi Chen, Xiao-Wei Zhou, Meimei Huo -, Bo Zhao, Minyi GuoInternational Journal of Smart Grid and Clean Energy 2013 · DOI

TL;DR

This paper proposes a novel short-term wind power forecasting method by combining Empirical Mode Decomposition (EMD) with Radial Basis Function Neural Networks (RBFNN). By decomposing complex wind power signals into simpler components, the method achieves higher prediction accuracy and adaptability compared to traditional direct prediction approaches.

What Problem Does This Paper Address?

Wind power generation is inherently unstable, intermittent, and uncontrollable, posing significant challenges for its integration into smart grids. Accurate short-term forecasting of wind power is crucial for grid stability and efficient energy management. Traditional forecasting methods often struggle with the non-linear and non-stationary characteristics of wind power data, leading to suboptimal prediction accuracy. This paper aims to improve forecasting performance by addressing these complexities.

Methods

The core of the proposed method lies in a two-stage approach:

  1. Empirical Mode Decomposition (EMD): The raw wind power signal is decomposed into several Intrinsic Mode Functions (IMFs) and a residue. EMD is a data-driven technique that breaks down a signal into its constituent oscillatory modes, each representing a different characteristic scale. This process effectively separates the signal into components that are easier to analyze and predict, isolating the influence of non-linear and non-stationary behaviors.
  2. Radial Basis Function Neural Networks (RBFNN): For each decomposed IMF and the residue, an independent RBFNN model is trained. The input variables for these RBFNNs include wind speed, wind direction, and historical wind power data. The RBFNN is chosen for its ability to approximate complex functions and its suitability for classification and forecasting tasks.
  3. Aggregation: The individual prediction results from each IMF and the residue are then aggregated to produce the final short-term wind power forecast.

Datasets

The experimental data used in this paper is from the Zhejiang Provincial Electric Power Test Research Institute. The dataset consists of wind power output sampled every 60 minutes. For analysis, 974 data points were used. The paper mentions using 900 samples for training and the remaining 74 samples for testing the models.

Key Insights and Ideas

  • Decomposition for Simplicity: The key insight is that by decomposing a complex, non-stationary signal like wind power into simpler, more regular components (IMFs and residue), prediction becomes significantly easier and more accurate. This is analogous to breaking down a complex problem into smaller, manageable sub-problems.
  • Isolation of Influences: EMD effectively isolates the influence of different oscillatory modes, reducing the "mutual influence" between them. This isolation helps mitigate the impact of non-linear and non-stationary behaviors on the overall prediction.
  • Tailored Models: Training separate RBFNN models for each IMF and the residue allows for specialized prediction based on the specific characteristics of each component. This is more effective than a single model trying to capture all the signal's complexities.

Key Results

The paper demonstrates that the proposed EMD-RBFNN method achieves superior prediction accuracy compared to a direct prediction model using only RBFNN. Specifically, the EMD-RBFNN method resulted in:

  • Lower Mean Absolute Error (MAE): 22.075 kW (vs. 34.905 kW for direct RBFNN).
  • Lower Mean Square Error (MSE): 2.485 (vs. 4.293 for direct RBFNN).
  • Lower Mean Absolute Percent Error (MAPE): 17.582% (vs. 29.543% for direct RBFNN).
  • Lower Mean Square Percent Error (MSPE): 8.742% (vs. 18.265% for direct RBFNN).

These quantitative results indicate that the EMD-RBFNN approach provides a more precise and reliable forecast.

Limitations and Future Work

  • IMF Selection: While the paper mentions excluding IMF1 (high-frequency component) from the forecast, the rationale for excluding specific IMFs or the residue could be further elaborated. The paper implies that IMF1 represents random interference, but a more rigorous justification might be beneficial.
  • Generalizability: The study was conducted on data from a single wind turbine. Further research could explore the applicability of this method to multiple wind turbines or entire wind farms, which might exhibit different characteristics.
  • Computational Cost: While not explicitly detailed, EMD can be computationally intensive. Future work could investigate optimizing the decomposition process or exploring alternative decomposition techniques that offer a better trade-off between accuracy and computational efficiency.
  • Parameter Optimization: The paper mentions using a literature-based method for determining hidden layer neurons in RBFNN. Further exploration into advanced hyperparameter tuning techniques for both EMD and RBFNN could potentially yield even better results.

Summary updated

Abstract

In order to effectively predict wind farm power with non-linear and non-stationary characteristics, a prediction model based on empirical mode decomposition (EMD) and radial basis function neural networks(RBFNN) was designed. The forecast model uses EMD to decompose the wind power into several intrinsic mode functions (IMF) and one residue. The RBFNN was used to construct a prediction model for each IMF component and the residue, the input variables of the prediction model are triple: wind speed, wind direction, and history wind power. All the prediction results of components were aggregated to obtain the ultimate prediction result. The simulation results show that compared with the traditional prediction method based on artificial neural networks, this method has high prediction precision and strong adaptability.
BibTeX
@article{zheng2013short,
  author = {Zengwei Zheng and Yuanyi Chen and Xiao-Wei Zhou and Meimei Huo - and Bo Zhao and Minyi Guo},
  title = {{Short-Term Wind Power Forecasting Using Empirical Mode Decomposition and RBFNN}},
  journal = {International Journal of Smart Grid and Clean Energy},
  year = {2013},
  doi = {10.12720/sgce.2.2.192-199},
}
Figure