一种利用多任务学习的短期住宅负荷预测方案

作为信息物理社会系统的一种具体形式，智能电网中的负荷预测，尤其是单个电力客户的短期负荷预测，在智能电力系统的规划和运营中将扮演越来越重要的角色.考虑到同一住宅小区用户之间的负荷行为的相似性，受多任务学习的启发，提出了一种基于多任务学习的有效住宅负荷预测方案.首先，利用K-means聚类技术和皮尔逊相关系数挑选出2个相似用户，进而将2个用户的负荷数据合并输入，并将双向长短时记忆网络作为共享层全面捕获2个用户数据之间的关系，然后送入2个全连接的任务相关的输出层.在真实的数据集上，将所提方案与几种典型的负荷预测方案进行全面比较.实验结果表明，与已有的深度学习预测方案相比，提出的多任务负荷预测方案提高了预测准确程度.

An Short-Term Residential Load Forecasting Scheme Using Multi-Task Learning

In smart grid regarded as specific embodying of cyber-physical-social system, load forecasting, especially short-term load forecasting for individual electric customers plays an increasingly role in planning and operation of smart power system. Considering the similarity of electricity consumption between users, inspired by multi-task learning, the article puts forward an effective residential load forecasting based on multi-task learning model. In detail, the K-means clustering technology and Pearson correlation coefficient are used to select two similar users. Then these two user's load data are merged as input, the bidirectional long short-term memory network is used as a sharing layer to fully capture the relationship between the data of the two users, and then two fully-connection task-specific output layers are respectively built. Based on real datasets, the proposed scheme is thoroughly compared with several typical deep learning based load forecasting schemes. Experiments show that proposed multi-task learning scheme improves the prediction accuracy compared with the existing deep learning prediction scheme.