Abstract: Developing renewable energy generation is a crucial measure in achieving the strategic goals of “carbon peaking and carbon neutrality.” The construction of a new power system dominated by renewable energy demands higher flexibility from the power grid. As a typical representative of renewable energy, wind power has become an integral component of the global power system, owing to its abundant resources, minimal environmental impact and well-established industrial foundation. Wind power ramp events refer to the sharp fluctuations in wind power output over short time intervals. With the continuous growth in wind power capacity and its increasing share in the energy mix, these ramp events have emerged as a significant risk factor affecting the economic and reliable operation of power grids. The potential impact on grid stability is substantial. However, wind power ramp events exhibit complex characteristics, including transience, uncertainty and non-linearity, which pose significant challenges for accurate prediction. Reliable forecasting of wind power ramp events is essential for facilitating wind power integration, guiding rational grid dispatch, and mitigating the risks of power imbalances due to wind power grid integration.
This paper proposes an online prediction framework for wind power ramp events that adapts to concept drift, which consists of an online prediction module, a trend compression module and a ramp identification module. First, based on the adaptive random forest model, the wind power output online prediction module is constructed, which uses an online ensemble learning method to adaptively update the model parameters, adapt to the concept drift in wind power output, and achieve accurate power prediction. Second, based on the swinging door algorithm optimized by the Grey Wolf Optimization (GWO), the trend compression module is constructed, which calculates the optimal tolerance coefficient and compresses the power prediction data, extracts the significant trend, and reduces the interference of noise data on ramp prediction. Finally, based on the fluctuation trend-based ramp identification algorithm, the ramp identification module is constructed, which avoids the multiple and false detection caused by complex local fluctuations, and identifies the ramp events to obtain the final ramp prediction results. The integration of these modules ensures both the accuracy and robustness of the wind power ramp event predictions.
In the empirical study, this paper utilizes wind power data from two wind farms located in the Fujian and Zhejiang provinces in China to validate the proposed framework. Additionally, three error metrics and three accuracy metrics are employed to evaluate the prediction performance of the framework. The results indicate that the framework outperforms several benchmark models in terms of both prediction accuracy and robustness. Notably, the adaptive random forest model effectively addresses concept drift in the wind power data, enhancing the model’s generalization ability. Moreover, the GWO algorithm optimizes the parameters of the swinging door algorithm, determining the optimal tolerance coefficient for different wind farm data, which aligns the compression results more closely with the fluctuations in wind power. The fluctuation trend-based ramp identification method mitigates multiple and false detections caused by minor local fluctuations in wind power, accurately capturing various types of ramp events.
In summary, the online prediction framework for wind power ramp events, which adapts to concept drift, demonstrates strong predictive performance. This framework holds significant potential for optimizing the operational control and scheduling strategies of wind farms, thereby enhancing the economic efficiency and reliability of wind power. Additionally, its ability to handle real-time data fluctuations makes it suitable for practical applications in diverse wind farm environments. Based on this framework, future research could integrate wind power data into numerical weather prediction data to improve ramp events predictions, further enhancing the model’s predictive capabilities.

Key words: wind power, ramp events, concept drift, online prediction

摘要： 风电功率爬坡事件是由风速、风向等气象因素引发的风电功率短时剧烈波动,容易导致系统功率失衡,对电网的电能质量和安全运行带来重大威胁。概念漂移是指数据分布或生成过程随时间变化的现象,气象环境及风电机组运行工况的变化可能引发风电功率的概念漂移,从而降低对风电爬坡事件预测的有效性和可靠性。为此,本文提出了一种适应概念漂移的风电功率爬坡事件在线预测框架,由三个模块组成。首先,基于自适应随机森林模型构建风电功率在线预测模块,利用在线集成学习方法自适应更新模型参数,以适应风电功率中的概念漂移并实现准确的功率预测。其次,利用灰狼算法优化的旋转门算法构建趋势压缩模块,对功率预测数据进行压缩,提取功率变化中的显著趋势,减少噪声对爬坡识别的干扰。最后,基于波动趋势构建爬坡识别模块,避免复杂局部波动导致的误检和漏检,得到准确的爬坡预测结果。实验结果表明,所提出的在线预测框架在预测精度和鲁棒性上均优于所有对比模型,能够为电力系统的安全调度提供可靠参考。

关键词: 风电功率, 爬坡事件, 概念漂移, 在线预测