郭军红, 王小萱, 汪月新, 李薇, 丁一, 贾宏涛. Copula分位数回归方法在风电超短期出力预测上的应用[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2023.12.15.003
引用本文: 郭军红, 王小萱, 汪月新, 李薇, 丁一, 贾宏涛. Copula分位数回归方法在风电超短期出力预测上的应用[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2023.12.15.003
GUO Junhong, WANG Xiaoxuan, WANG Yuexin, LI Wei, DING Yi, JIA Hongtao. Enhancing ultra-short-term wind power forecasting using the Copula quantile regression method[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2023.12.15.003
Citation: GUO Junhong, WANG Xiaoxuan, WANG Yuexin, LI Wei, DING Yi, JIA Hongtao. Enhancing ultra-short-term wind power forecasting using the Copula quantile regression method[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2023.12.15.003

Copula分位数回归方法在风电超短期出力预测上的应用

Enhancing ultra-short-term wind power forecasting using the Copula quantile regression method

  • 摘要: 风电出力具有较强的随机性和波动性,相比于传统预测,分位数预测方法能够提供全面的风电功率概率分布信息,可实现更可靠的风电出力预报,对电网系统的安全和稳定运行具有重要意义. 以甘肃某风电站为案例,将数据按8:2划分为训练集和测试集,采用基于Copula的分位数回归方法(QCopula)进行功率区间预测,并与三个传统的分位数回归方法进行比较. 结果显示,在不同置信区间下QCopula的修正预测区间精度范围在0.701~0.773之间,预测精度平均值比传统分位数回归(QR)、随机森林分位数回归(QRF)和长短期记忆神经网络分位数回归(QLSTM)分别高出15%、9%和13%,优于其他三种分位数预测方法. 分位数交叉验证中,QCopula未出现分位数交叉,每个样本点的功率预测值均随概率值单调递增,而QR、QRF、QLSTM均出现不同程度的分位数交叉现象. 综上所述,QCopula可以表征更小的区间宽度和更高的区间覆盖率,且分位数曲线不存在交叉,可信度较高.

     

    Abstract: In recent years, the shift toward renewable energy in China’s power industry has been remarkable, with the installed capacity of renewables surpassing that of coal-fired power. Among these, wind power output plays a pivotal role, although it is characterized by its strong randomness and volatility. Traditional prediction methods fall short as they cannot provide comprehensive probability distribution information on the wind power output. To bridge this gap, quantile prediction methods have emerged as superior options for achieving reliable wind power output predictions, which are crucial for the safe and stable operation of power grid systems. To address the inherent unpredictability of wind power, this study introduces a quantile regression method based on Copula (QCopula). The Copula function captures the correlation between the marginal distribution functions of the random variables and their joint distribution function. The process begins with selecting an optimal Copula function using the Akaike Information Criterion (AIC). This function elucidates the relationship between wind power and wind speed, enabling the expression of the conditional probability distribution function of power. By considering different conditional probability values, we obtained wind power prediction results at different quantiles, leading to interval prediction results across different confidence intervals. These results were compared with three traditional quantile regression methods (Quantile Regression (QR), Quantile regression Random Forests (QRF), and Quantile regression Long Short-Term Memory (QLSTM)) using three elevation metrics: Predictive Interval Coverage Probability (PICP), Predictive Interval Normalized Average Width (PINAW), and Corrected Predictive Interval Accuracy (CPIA). This comparison was aimed at evaluating the interval prediction accuracy of the four quantile regression methods. Finally, the crossover of the quantile curves for each method was analyzed. A case study was conducted at a wind power plant in Gansu Province, utilizing wind speed and power data (measured in MW at 15-minute intervals) from September 2022 to June 2023. With 29088 sample points in total, the data were divided into training and testing sets in an 8:2 ratio. The training set facilitated model development through various quantile regression methods, whereas the testing was used to evaluate the accuracy of each model. The results showed that the QCopula consistently outperformed the other methods across different confidence intervals, with its modified prediction interval accuracy ranging between 0.701 and 0.773. On average, it exceeded QR, QRF, and QLSTM by 15%, 9%, and 13%, respectively. Notably, the QCopula maintained a consistent increase in the predicted power values for each sample point with probability, without any instances of quantile crossing, a common issue observed in QR, QRF, and QLSTM. In summary, the QCopula offers narrower interval widths and higher interval coverage without the drawback of quantile curve crossing, thereby ensuring higher reliability.

     

/

返回文章
返回