Hierarchical dynamics-aware deep learning for electricity consumption prediction in discrete manufacturing enterprises

Predicting the overall electricity consumption of manufacturing enterprises is an important step in achieving energy demand-side management and industrial energy system stability. Due to the diversity of consuming units and complexity of dynamic consuming distribution in discrete manufacturing enterprises (DME), their electricity consumption exhibits hierarchical dynamics. However, current temporal modeling and analysis methods cannot overcome the coupled variation caused by hierarchical dynamics, which makes accurate electricity consumption prediction (ECP) in DME still a bottleneck problem. In this paper, a hierarchical dynamics sensitive prediction network (HDSPN) is proposed to realize the ECP in DME. First, the basic structure and distribution of electricity consumption in DME are demonstrated to analyze the hierarchical dynamics caused by discrete production characteristics. Second, the HDSPN is designed to perform frequency recognition and two-dimensional reorganization on one-dimensional temporal data to extract the inter- and intra-period dynamic information of electricity consumption. Through the integration training using data from the overall enterprise and its internal multiple consuming units, HDSPN can sensitively capture the hierarchical dynamics about the electricity consumption of DME. Finally, verified on a typical DME, Dongfang Electric Machinery Co., Ltd., the root mean square error (RMSE) of HDSPN reaches the optimal 0.020194, 0.056044, 0.114503 respectively in very short-term, short-term and long-term prediction compared with the mainstream models, thereby meeting the multi-timescale requirements of demand-side management.