基于多任务均衡学习的建筑多元柔性负荷预测方法研究
Research on multivariate flexible load prediction method for buildings based on multi-task balanced learning
摘要：
建筑柔性负荷调度是缓解新型电力系统供需矛盾的有效手段,但造成建筑负荷更为复杂多变,并且空调与电力负荷的相互耦合使得建筑多元负荷的准确预测更为困难。为解决上述问题,本文通过采用结合分时电价的聚类分析方法识别建筑能源系统的柔性用能特征,优化预测模型输入参数,基于长短期记忆(LSTM)神经网络的时序记忆功能和梯度归一化多任务学习(GNMTL)的耦合信息共享机制,构建了建筑多元负荷预测模型,以提高模型对多任务的均衡学习能力。以北京某办公建筑为案例,验证了该方法的有效性。结果表明,本文所提出的方法可以兼顾对建筑负荷柔性特征及多元耦合关系的学习,相较于现有方法R2提高2.2%～11.4%,RMSE和MAE分别降低6.4%～43.6%、24.9%～55.5%。本文研究为建筑柔性负荷调度提供了更为准确的负荷预测解决方案。
Abstract：
Building flexible load scheduling is an effective means to alleviate the contradiction between supply and demand of the new power system, but it causes building loads to be more complex and variable, and the mutual coupling of air conditioning and power loads makes it more difficult to accurately predict building multivariate loads. In order to solve the above problems, this paper constructs a building multivariate load prediction model by using a cluster analysis method combined with time-of-use tariffs to identify the flexible energy use characteristics of the building energy system and optimizing the input parameters of the prediction model based on the temporal memory function of the long short term memory (LSTM) neural network and the coupled information sharing mechanism of the gradient-normalized multi-task learning (GNMTL) to improve the model’s multi-task balanced learning. An office building in Beijing is used as a case study to verify the effectiveness of the method. The results show that the proposed method can take into account the learning of the flexible characteristics and multivariate coupling relationship of building loads, with R2improved by 2.2% to 11.4% and RMSE and MAE reduced by 6.4% to 43.6% and 24.9% to 55.5%, respectively, compared with the existing methods. The research in this paper provides a more accurate load forecasting solution for building flexible load scheduling.
Keywords:multi-task balanced learning; flexible energy use feature recognition; multivariate load prediction; long short term memory (LSTM) neural network; K-means cluster analysis