基于门控循环单元网络与模型融合的负荷聚合体预测方法

随着智能电表的普及,以智能电表数据为基础,可按需求灵活划分不同规模的负荷聚合体并开展预测。由于负荷聚合体规模差异较大,并与用户负荷特性关系密切,传统预测方法不再适用。为此,提出了一种基于门控循环单元(GRU)网络与模型融合的负荷聚合体预测方法。首先,通过分布式谱聚类算法获得负荷特性相近的负荷群体,然后进行分组预测,采用GRU作为元模型,对时间序列进行动态建模,利用随机森林算法融合多个结构不同的GRU网络,实现对负荷群体的预测,最终将各群体预测值求和得到负荷聚合体预测值。算例表明,得益于分组预测、动态时间建模及模型融合技术,所述方法能充分利用不同模型的结构优势,发现时间序列动态规律,在不同时间尺度下预测精度更高,对不同规模的负荷聚合体适用性更强。

Aggregated Load Forecasting Method Based on Gated Recurrent Unit Networks and Model Fusion

With the popularity of smart meters, the aggregated load can be divided flexibly into different sizes according to different requirements and be predicted based on the measurement data. Due to the large difference in the scales of aggregated loads and the close relationship with the load characteristics of users, the traditional prediction method is no longer applicable. This paper proposes an aggregated load forecasting method based on gated recurrent unit(GRU)networks and model fusion. Firstly, load groups with similar load characteristics are clustered by the distributed spectral clustering algorithm, then grouping predictions are employed. Secondly, GRU is adopted as a meta-model to perform dynamic modeling of time series, and several different structures of GRU networks are fused by random forest algorithm to realize the load group forecast. Finally, the aggregated load forecast value can be obtained by summing prediction value of each group. Benefiting from the grouping prediction, dynamic time modeling, and model fusion technology, the proposed method can make full use of advantages of different model structures and discover the dynamic rule of time series. The proposed method achieves higher prediction accuracy and higher adaptability for aggregated load with different scales.