基于SWLSTM-Stacking集成学习的源荷区间预测方法

A Source-Load Interval Prediction Method Based on SWLSTM-Stacking Integrated Learning

针对光伏出力和电动汽车充电特性的随机特性对电力系统的冲击不断增强,准确及时的源荷预测是实现增强电力系统适应性和稳定性的重要课题。因此,提出一种基于共享权重长短期记忆网络(shared weight long short-term networks,SWLSTM)与Stacking集成模型相结合的源荷区间预测方法。首先,光伏出力存在时序性特征,采用局部线性嵌入改进k-means算法聚类提取特征日,在实现数据降维同时,减少了网络训练难度;其次,在Stacking集成模型的框架下,将SWLSTM作为元学习器,并通过Q统计量筛选合适的基学习器模型,从而实现多模型融合的多异学习器Stacking集成学习的源荷预测;紧接着,为了得到预测的不确定信息,引入置信度区间预测;最后,采用实测数据对本文所提方法进行验证。结果表明改进k-means算法能够降低其求解难度,加快求解速度,可以快速获取聚类特征;所引入集成学习模型和置信度区间,有效表征源荷预测的不确定性,提升区间预测模型的泛化能力。

In response to the increasing impact of the stochastic characteristics of photovoltaic(PV)output and electric vehicle(EV)charging characteristics on the power system,accurate and timely source-load prediction is becoming even more important to realize the enhancement of the adaptability and stability of the power system.Therefore,this paper proposes a source-load interval prediction method based on the combination of shared weight long short-term networks(SWLSTM)and Stacking integration model.Firstly,as the PV outflow has temporal characteristics,the local linear embedding improved k-means algorithm is used to cluster and extract the feature days,which reduces the network training difficulty while realizing the data dimension reduction.Secondly,under the framework of Stacking integrated model,SWLSTM is used as a meta-learners,and suitable base-learners models are screened by the Q-statistic,which realizes the source-load prediction of the multi-differential learning ware Stacking integrated learning of the multi-model fusion.Immediately after that,confidence interval prediction is introduced in order to get the uncertain information of prediction.finally,the method proposed in this paper is validated using real data,and the results show that the improved k-means algorithm can reduce its solving difficulty,speed up the solving speed,and can quickly obtain the clustering features,the introduced integrated learning model and confidence interval can effectively characterize the uncertainty of the source load prediction,and improve the generalisation ability of the interval prediction model.