摘要
针对短期电力负荷随时间变化呈现随机性和不确定性问题,提出了一种基于加权灰色关联投影算法Bagging-Blending的融合模型。首先,采用加权灰色关联投影算法对电力负荷中各影响因素(如天气、温度、湿度、日期类型等)进行分析,以选取历史负荷特征。在此基础上,分别将各单一模型SVR(support vector regression)、KNN(K-nearest neighbor)、GRU(gate recurrent unit)、XGBoost(eXtreme Gradient Boosting)、LightGBM(light gradient boosting machine)、CatBoost(Categorical features gradient Boosting)嵌入Bagging集成算法中以提升模型的稳定性和泛化能力。同时利用Pearson相关系数对各单一模型进行相关性分析。然后,依据模型对数据观测空间角度的不同,使用Blending模型对相关性小的模型进行融合。最后,通过新英格兰地区电力负荷数据ISO New England进行验证。所提融合模型与传统单模型(SVR、GRU)和其他融合模型(Bagging-XGBoost、最优加权的GRU-XGBoost)相比,具有较强的泛化能力和较高的稳定性与预测精度。
In order to solve the problems of randomness and uncertainty of short-term power load with time,we propose a fusion model based on weighted grey relational projection algorithm Bagging-Blending.First,we use the weighted grey relational projection algorithm to analyze the influencing factors of power load(such as weather,temperature,humidity,date type,etc.) in order to select the characteristics of historical load.On this basis,we embed each single model in the Bagging integration algorithm,including support vector regression(SVR),K-nearest neighbor(KNN),gate recurrent unit(GRU),e Xtreme Gradient Boosting(XGBoost),Light Gradient Boosting Machine(Light GBM) and Categorical features gradient Boosting(Cat Boost),to improve the stability and generalization ability of the model.At the same time,we use the Pearson correlation coefficient to analyse the correlation of each single model.Second,according to the different angles of the data observation space of the model,we use the Blending model to fuse the models with low correlation.Finally,we verify the model by the power load data of in New England,ISO New England.Compared with the traditional single model(SVR,GRU)and other fusion models(Bagging-XGBoost,optimal weighted GRU-XGBoost),the proposed fusion model has stronger generalization ability,higher stability and prediction accuracy.
作者
徐耀松
王德兴
XU Yaosong;WANG Dexing(Faculty of Electrical and Engineering Control,Liaoning Technical University,Huludao 125105,China)
出处
《信息与控制》
CSCD
北大核心
2022年第2期188-200,共13页
Information and Control
基金
国家自然科学基金(51974151,71771111)。
