TL;DR
This paper introduces a novel wind power interval prediction method that combines Phase Space Reconstruction (PSR), Broad Learning System (BLS), and Quantile Regression (QR) with an adaptive rolling error correction mechanism. This approach aims to improve the accuracy and reliability of predicting wind power output ranges, crucial for grid stability and planning, by effectively handling the inherent uncertainty and volatility of wind energy.
What Problem Does This Paper Address?
Wind power is a vital renewable energy source, but its intermittent and unpredictable nature poses significant challenges for grid integration and management. Accurate prediction of wind power output is essential for grid security and stability. While point prediction offers a single value, interval prediction provides a range, offering more comprehensive information for decision-makers. However, traditional interval prediction methods often suffer from low accuracy, particularly in capturing the complex, non-linear dynamics of wind power. This paper addresses this gap by proposing a method that not only models the underlying data patterns but also actively corrects prediction errors in a dynamic, adaptive manner.
Methods
The proposed method integrates several key techniques:
- Phase Space Reconstruction (PSR): This technique is used to transform the one-dimensional wind power time series into a higher-dimensional space. This process helps to uncover hidden correlations and patterns within the data that might not be apparent in the original one-dimensional representation. The optimal embedding dimension and delay time are determined using the C-C phase method.
- Broad Learning System (BLS): BLS is employed as the core prediction model. Unlike traditional deep neural networks, BLS offers a simpler structure with faster training times. It utilizes a feature layer and an enhancement layer to learn the complex relationships between input and output data. The paper highlights BLS's advantage in reducing parameter tuning and computational cost.
- Quantile Regression (QR): After obtaining point predictions from BLS, QR is used to construct the initial prediction intervals. QR is chosen because it does not assume a specific data distribution and can directly estimate conditional quantiles, which are essential for defining interval boundaries.
- Adaptive Rolling Error Correction: This is the novel contribution of the paper. Instead of a static error correction, this method dynamically corrects errors based on the distribution of prediction errors within specific power interval segments (e.g., 0-200 kW, 200-400 kW). Non-parametric kernel density estimation is used to model these error distributions, and an optimal correction index is employed to determine the best correction power for each segment and boundary (upper/lower). This rolling correction mechanism continuously updates predictions as new data becomes available.
Datasets
The paper utilizes three wind power datasets from different regions to validate the proposed method:
- Set 1: Jiangsu Province, China. Rated power 3.5 MW. Data from January 1, 2020, to April 30, 2020 (121 days), with a 15-minute sampling interval.
- Set 2: Hubei Province, China. Wind turbine model EN-141/2.5, rated power 2.5 MW. Data from January 1, 2021, to December 31, 2021 (365 days), with a 10-minute sampling interval.
- Set 3: Offshore wind farm in Belgium. Data collected every 15 minutes from January 1, 2020, to December 31, 2020 (366 days).
The paper also discusses data preprocessing steps, including handling negative values and ensuring data normality.
Key Insights and Ideas
- Uncovering Hidden Dynamics: PSR is crucial for transforming the seemingly random one-dimensional wind power data into a higher-dimensional space where underlying temporal dependencies become apparent. This is a foundational step for more accurate modeling.
- Efficiency of BLS: The paper strongly advocates for BLS due to its computational efficiency and reduced need for extensive hyperparameter tuning compared to traditional deep learning models. This is particularly important for real-time prediction applications.
- Segmented Error Correction: The most significant insight is the realization that prediction errors are not uniform across all power output levels. By segmenting the power output and applying specific error correction strategies to each segment, the method achieves a more nuanced and effective correction, leading to improved accuracy.
- Adaptive Nature: The "rolling" aspect of the error correction is key. It implies that the model continuously learns and adapts to new error patterns, making it robust to changing wind conditions.
Key Results
The proposed PSR-BLS-QR with adaptive rolling error correction demonstrates significant improvements over traditional methods:
- Improved Accuracy: The method achieves an average prediction accuracy improvement of about 6-14% compared to traditional interval prediction methods.
- Narrower Interval Width: The proposed method achieves this accuracy improvement with a narrower interval width, indicating more precise predictions.
- Reduced Running Time: BLS contributes to a significantly lower running time for both training and testing compared to other models like BPNN, LSTM, and RBF.
- Enhanced Coverage Probability: The Prediction Interval Coverage Probability (PICP) is consistently higher across different confidence intervals and datasets after applying the error correction. For example, in Set 1, the average PICP improved by approximately 8% after correction.
Limitations and Future Work
- Data Preprocessing: While the paper mentions data cleaning, the impact of different preprocessing techniques on the final results could be further explored.
- Generalizability of Segmentation: The choice of 200 kW as the segmentation point for error correction is empirical. Further research could investigate optimal segmentation strategies for different datasets or conditions.
- Computational Cost of Error Distribution: While BLS is efficient, the kernel density estimation for error distributions across multiple segments and confidence levels might still be computationally intensive for very large datasets or real-time applications requiring extremely low latency.
- Future Work: The authors suggest exploring other methods for adaptive error correction and potentially integrating more advanced feature engineering techniques to further enhance prediction accuracy. The paper also implies that the method's effectiveness could be tested on other renewable energy sources.
Summary updated