基于可解释性优化堆叠模型的风电功率预测  

WIND POWER FORECASTING BASED ON INTERPRETABLE OPTIMAL STACKING MODEL

作　　者：戚焕兴 卓毅鑫 李凌 殷林飞 秦意茗 蒙文川 Qi Huanxing;Zhuo Yixin;Li Ling;Yin Linfei;Qin Yiming;Meng Wenchuan(Guangxi Power Grid Dispatching Control Center,Nanning 530023,China;Beihai Power Supply Bureau,Guangxi Power Grid,Beihai 536000,China;School of Electrical Engineering,Guangxi University,Nanning 530004,China;Energy Development Research Institute,China Southern Power Grid,Guangzhou 511458,China)

机构地区：[1]广西电网有限责任公司电力调度控制中心,南宁530023 [2]广西电网有限责任公司北海供电局,北海536000 [3]广西大学电气工程学院,南宁530004 [4]南方电网能源发展研究院有限责任公司,广州511458

出　　处：《太阳能学报》2025年第4期559-569,共11页Acta Energiae Solaris Sinica

基　　金：广西电网公司科技项目(046000KK52220007)。

摘　　要：鉴于当前风电功率预测算法在堆叠建模过程所存在试错性与弱解释性问题,提出一类应用于风电功率预测的可解释性优化堆叠模型(interpretable optimal stacking model,IOSM)。首先,建立初始冗余基学习器并进行特征权重解释;其次,构建一个综合衡量模型性能与计算代价的特征优化度量,优化关键特征并训练形成优化后的基学习器;最终,根据特征优化度量对基学习器进行优化选择,搭建得到元学习器。经如上步骤,即完成搭建IOSM。所提算法在广西典型风电场的算例表明,IOSM相对于最优单一模型的RMSE与MAE指标上分别降低13.01%和18.23%;相对于其他各类主流组合预测算法,该文算法在RMSE与MAE指标分别降低8.24%与10.28%的同时,至少降低了142.38%的建模计算代价。所提算法的有效性及先进性得到验证,为风电功率预测的可解释性优化建模上提供了新的思路与方法。High accuracy of wind power forecasting is an important support to ensure the balance of the power and load and stable operation of power system.In view of the trial-error and weak interpretative problems of current wind power forecasting algorithms in stack modeling,an interpretable optimal stacking model(IOSM)is proposed in this paper,which is applied to wind power forecasting.Firstly,establish an initial redundant base learner and interpret the feature weights.Further,construct a feature optimization metric that comprehensively measures model performance and computation cost,to optimize and select key features and train them to form optimized base learners.Finally,based on feature optimization metrics,the meta learner is established by the the optimized base learner.As above,the complete IOSM is constructed.The proposed algorithm is applied to typical wind farms in Guangxi.Compared with the optimal single model,the RMSE and MAE indexes of IOSM are reduced by 13.01%and 18.23%respectively.Compared with other mainstream combinatorial algorithms,the RMSE and MAE indexes of IOSM can be decreased by 8.24%and 10.28%respectively,while still reducing 142.38%modeling computation cost at least.The effectiveness and advancement of the proposed algorithm are verified,which provides a new idea and method for interpretability modeling of wind power forecasting.

关 键 词：风电功率预测 可解释性 堆叠模型 特征优化 集成学习 

分 类 号：TM614[电气工程—电力系统及自动化]