Sleep-induced Network with Reducing Information Loss for Short-term Load Forecasting

Short-term load forecasting (STLF) plays an important role in real-time decision-making and management of the power system while is still a challenging task. Considering that sleep improves brain memories and cognitive processes, this paper explores a approach of integrating biological mechanisms to reduce information loss of networks, and hence proposes a sleep-induced network (SI-Net) by analogy for achieving high-performance STLF. Firstly, through mimicking the sleep process, a multi-level bionic flowchart of the SI-Net is designed to integrate the gated, attention, parallel, cooperative, and asynchronous mechanisms, which not only encode features from coarse to fine but also enhance the fitting capability at the feature layer. Secondly, through imitating the brain memory paths during sleep, the primary and secondary memory paths are designed to update and store information, respectively, and their independence and collaboration avoid information loss in the SI-Net. Thirdly, the loss function constructed by the Gaussian kernel makes nonlinear errors linearly separable in the high-dimensional space, being beneficial to train the SI-Net. The experiments with real-world load datasets are performed and the results show that the SI-Net outperforms 15 baselines and presents high accuracy and stability. Bionically-inspired ideas are promising to design high-performance forecasting networks for energy systems.