基于XGBoost-Bin自动功率极限计算的风电机组健康性能评估及预测

Health performance evaluation and prediction methods for wind turbines based on XGBoost-Bin automatic power limit calculation

针对数据量较大情形下风电机组健康性能评估区域难划分、健康性能预测精度低等问题,提出基于XGBoost-Bin的自动功率曲线极限算法,建立考虑多数据特征的机组健康性能预测模型。提出一种基于XGBoost-Bin的曲线构建算法,获取表征风电机组运行状态的静态最优功率曲线;提出一种改进自动功率曲线极限计算的风电机组健康性能评估方法,以准确划分机组健康性能评估区域,并通过综合分析实际与理论静态最优功率曲线偏差、发电效能等指标评估机组健康性能;提出基于多数据特征广义回归神经网络的风电机组健康性能预测模型,以提升健康性能的预测精度。最后,以内蒙古塞罕坝风电场20台风电机组为例表明,与传统自动功率曲线极限计算方法相比,所提风电机组健康性能评估方法的评估准确度提升了0.1026,所提预测模型能够提升机组健康性能的预测精度,比传统随机森林预测模型R2提升了0.017。

The health performance evaluation areas of the wind turbines are difficult to divide thanks to the large data,which influences the health performance prediction accuracy.To tackle these problems,an automatic power curve limit calculation algorithm based on XGBoost-Bin was proposed,and a health performance prediction model considering multiple data characteristics was developed.A static optimal power curve construction algorithm based on XGBoost-Bin was proposed to reflect the operating states of the wind turbines.An improved automatic power curve limit calculation was proposed to accurately divide the wind turbine health performance evaluation areas,and evaluate the health performance by analyzing the deviation of the actual and theoretical static optimal power curves and power generation efficiency.A wind turbine health performance prediction model based on multi-data features Generalized Regression Neural Network(GRNN)was proposed to improve the accuracy of wind turbine health performance prediction accuracy.20 wind turbines of Saihanba wind farm in Inner Mongolia of China were taken as an example to be analyzed.The experimental results showed that compared with the traditional automatic power curve limit calculation methods,the evaluation accuracy of the proposed health performance evaluation method for wind turbines was improved by 0.1026.The proposed prediction model could enhance the prediction accuracy of the health performance of wind turbines.Compared with the traditional random forest prediction model,the indicator of R2 with the proposed prediction model was enhanced by 0.017.