A general energy-aware framework with multi-modal information and multi-task coordination for smart management towards net-zero emissions in energy system

Deep learning plays a crucial role in advancing the smart management of energy systems, contributing significantly to improving energy efficiency and operational security. However, the limited adaptability to multi-modal energy information and low coordination in multiple energy tasks lead to difficulties in making accurate decisions for energy management. To this end, we proposed a novel general energy-aware framework with multi-modal information and multi-task coordination for smart management in energy systems. An adaptive transformation method was presented for converting multi-modal data to a general format by reassigning feature positions based on similarities. The proposed energy-aware framework fed with the generalized multi-modal data, going through feature extraction via a progressive vision backbone, and then produced outputs for multiple energy tasks. The adaptive loss weighting method was proposed to coordinate convergence rates and magnitudes of losses among multiple energy-related tasks. A series of experiments were conducted in practical energy systems to validate technical feasibility of proposed energy-aware framework. The performance metrics for multiple energy tasks including predictive maintenance, energy prediction and control optimization were 0.994, 0.942 and 0.945, and their performances remain relatively stable at 0.990, 0.920 and 0.952 for multi-task learning. The model performances can be increased by 7.19 % through adopting adaptive transformation. Moreover, extensive comparative experiments demonstrated the proposed energy-aware model outperformed common machine learning and deep learning algorithms. Our study is expected to develop more general and flexible deep learning model for smart management to save energy and ensure security, thereby supporting the realization of net-zero emissions in energy systems.