Abstract
Wind power prediction is a critical technology for enhancing the stability and efficiency of power systems integrating renewable energy. This paper presents a combined forecasting model designed to mitigate the sensitivities of reconstructed parameters in chaotic time series analysis. By utilizing Phase Space Reconstruction (PSR) and an ensemble of neural networks, the model integrates information across different embedding dimensions. Experimental results from the Meijia wind farm demonstrate that a neural network-based combination of sub-models achieves a prediction error of less than 7%, significantly outperforming single-parameter and linear weighted models.
Introduction
As global energy systems transition toward sustainability, wind power has become a cornerstone of renewable energy portfolios. However, the inherent intermittency and randomness of wind speed present significant challenges for grid operators. Inaccurate forecasts can lead to grid instability or the costly maintenance of "spinning reserves"—backup power plants kept idling to compensate for sudden drops in wind output.
Traditional linear models often struggle to capture the complex, non-linear dynamics of wind. While chaos theory provides a framework for understanding such systems, its application is highly sensitive to the selection of reconstruction parameters, specifically the delay time ($\tau$) and the embedding dimension ($m$). This research addresses the limitations of single-parameter models by proposing a combined approach that leverages multiple perspectives of the system's chaotic structure.
Methodology: Chaos Verification and Reconstruction
The core of the proposed approach lies in Phase Space Reconstruction (PSR), based on Takens’ Theorem. This theorem posits that a single observed variable, such as wind power over time, contains sufficient information to reconstruct the multi-dimensional state of the entire underlying system.
Parameter Estimation
To reconstruct the phase space effectively, two parameters must be determined:
- Delay Time ($\tau$): Using the Mutual Information method, the researchers identified an optimal delay time of $\tau = 6$.
- Embedding Dimension ($m$): The Correlation Dimension method was employed to find the ideal dimension. By plotting the correlation integral against the scaling radius for various dimensions, the researchers determined that the system stabilizes at $m = 8$.
Proving Chaos
Before applying predictive models, it is essential to verify that the time series is indeed chaotic rather than purely random. The researchers calculated the Largest Lyapunov Exponent, which measures the rate of separation of infinitesimally close trajectories. A positive value (0.2548 in this study) mathematically confirms that the wind power series is chaotic, implying it is deterministic and predictable in the short term despite its apparent randomness.
The Combined Neural Network Model
Recognizing that any single embedding dimension might have "blind spots," the authors developed an ensemble approach. They constructed three separate Back-Propagation Neural Networks (BPNN), each using a different embedding dimension ($m=7, 8,$ and $9$). This allows the system to capture a broader range of the hidden dynamics.
Two methods of integration were tested:
- Linear Weighted Combination: This method uses quadratic programming to assign fixed weights to each sub-model based on their historical performance.
- Neural Network Combination: A second-tier "master" neural network is trained to synthesize the outputs of the three sub-models. This approach is superior because it can learn the non-linear relationships between the errors of the individual models.
Results and Discussion
The study utilized real-world data from a 3.0 MW turbine at the Meijia wind farm, sampled at 30-minute intervals. The data was filtered to the turbine's operational range (3 m/s to 15 m/s) to ensure relevance.
The proposed Neural Network Combination model emerged as the most accurate:
- Single Model ($m=8$): 6.90% Absolute Average Error.
- Linear Combination: 6.86% Absolute Average Error.
- Neural Network Combination: 6.70% Absolute Average Error.
As shown in the performance plot, the prediction values closely track the actual fluctuations of the wind power series. While the 0.16% improvement over the linear model may seem marginal, in large-scale power grid management, such gains translate into substantial cost savings and significantly reduced risks of grid instability.
Conclusion
By treating wind power as a chaotic time series and employing a multi-model ensemble, this research provides a robust solution to the sensitivity issues inherent in phase space reconstruction. The neural network combination effectively integrates information from different embedding dimensions, leading to higher forecast accuracy. Future work may involve integrating Numerical Weather Prediction (NWP) data to extend the reliable forecasting horizon beyond the short-term window addressed here.
Summary updated