基于注意力机制-集成学习耦合的电力时间序列数据补全方法

Electric Power Time Series Data Imputation Method Based on AttentionMechanism and Ensemble Learning

基于电力系统所采集的海量数据进行运行管控,在维持电网的安全、稳定和经济运行上发挥着重要的作用。由于设备故障和通信中断等原因,量测数据可能会出现缺失。因此,如何补全缺失数据成为电力系统研究的热点与关键问题。然而,单一的数据补全模型往往难以应对复杂多变的数据环境,如大段缺失和时间序列特征多样等,不同模型处理数据的能力各有优缺点。基于此,提出AttnStack算法,一个耦合注意力机制和集成学习的电力时间序列数据补全方法。该算法综合了先进模型图循环补全网络模型(GRIN)和时间序列双向循环补全模型(BRITS)的预测结果,引入时间断面的上下文信息,通过注意力机制在每一时间断面上动态调整每个个体学习器输出的权重。综合GRIN和BRITS算法使之具备同时处理剧烈波动和平缓稳定场景的能力。通过在来自爱尔兰的实际居民用户智能电表数据上进行实验,相比于仅使用单个模型,使用单一模型(3个GRIN)的集成算法均方误差最少降低了4.87%,耦合GRIN和BRITS算法平均绝对误差降幅高达19%,证明AttnStack算法汇聚了各模型的优势,比单一模型更能应对大波动、少规律的用户负荷功率数据补全任务。

Leveraging the massive data collected from power systems for operational control plays a vital role in maintaining the safety,stability,and economic operation of the power grid.Due to equipment failures,communication interruptions,and other reasons,measurement data may be missing,making the imputation of missing data a hotspot and key issue in power system research.However,single data imputation models tend to struggle coping with complex and variable data environments,such as large missing segments,diverse time series features,etc.The ability of different models to deal with the data varies from one to another.In light of this,the AttnStack algorithm is proposed,a power time series data imputation method that couples attention mechanisms with ensemble learning.This algorithm integrates the imputation results of advanced models GRIN and BRITS,introduces the context information of time sections,and dynamically adjusts the output weight of each base learner on each time section through the attention mechanism.Combining the GRIN and BRITS models enables the handling of scenarios with both severe fluctuations and smooth stability.Experiments on real residential user smart meter data from Ireland show that comparing to use only one modle,using an ensemble algorithm of single models(three GRINs)reduces the mean squared error at least 4.87%,and coupling GRIN and BRITS algorithms reduces the mean absolute error by up to 19%,which demonstrates that,by converging the strengths of various models,AttnStack is more capable of handling the data imputation tasks of user load power data with large fluctuations and few regularity than a single model.