多变量特征引导下的风电机组功率曲线建模研究

Research on Wind Turbine Power Curve Modeling under the Guidance of Multivariable Features

中文关键词:风力发电机组  功率曲线建模  深度学习  多变量建模

keywords:wind turbine  power curve modeling  deep learning  multivariable modeling

中文摘要:针对传统功率曲线模型难以处理多变量非线性耦合关系的问题，提出了一种融合卷积神经网络、双向长短期记忆网络和注意力机制的风电机组多变量功率曲线建模方法。首先，采用涡流-四分位法对监控与数据采集 （supervisory control and data acquisition，SCADA） 系统数据进行异常值清洗，并通过偏互信息法筛选出多个关键特征变量。其次，利用卷积神经网络提取空间特征，通过双向长短期记忆网络捕捉时序动态特性，再结合注意力机制实现特征权重的自适应分配，从而有效处理多维环境因素与功率间的耦合关系。再次，为提升模型泛化能力，引入鲸鱼优化算法对超参数进行协同优化，并构建了包含均方根误差、平均绝对误差和决定系数的多指标评价体系。最后，基于某风电场并网运行风电机组SCADA实测数据开展对比分析。实验结果表明，与现有主流方法相比，所提模型建模精度更高，决定系数达到99.02%。研究成果验证了该模型能够有效表征多变量交互作用下风电机组的功率响应特性，为功率曲线建模提供了新的技术路径。

Abstract:To address the challenge that traditional power curve models struggle to handle multivariate nonlinear coupling relationships， this paper proposes a multivariate power curve modeling method for wind turbines integrating convolutional neural networks （CNN）， bidirectional long short-term memory （BiLSTM）， and an attention mechanism. Firstly， anomaly cleaning is performed on supervisory control and data acquisition （SCADA） data using the vortex-quartile method， and multiple key feature variables are identified via partial mutual information analysis. Secondly， CNN is employed to extract spatial features， BiLSTM captures temporal dynamic characteristics， and the attention mechanism enables adaptive allocation of feature weights—effectively resolving the coupling relationship between multidimensional environmental factors and power output. Thirdly， to enhance the model’s generalization performance， the whale optimization algorithm is introduced for collaborative hyperparameter optimization， and a multi-metric evaluation system is constructed， including root mean square error， mean absolute error， and coefficient of determination（R2）. Finally， experimental validation and comparative analysis are conducted based on measured SCADA data from grid-connected wind turbine generator systems in a wind farm. The results demonstrate that the proposed model achieves higher modeling accuracy compared with existing mainstream methods， with an R2 of 99.02%. This research confirms that the model can effectively characterize the power response characteristics of WTGS under multivariate interactions， providing a new technical pathway for power curve modeling.