Intelligent Modeling for Deep Peak Shaving Oriented Thermal Power-Molten Salt Heat Storage Coupled System by Improved Mamba-2

Deep and rapid peak shaving requirements of thermal power unit are becoming more significant in the low-carbon power system consisting of high-proportion renewable energy sources. In order to meet these requirements, this article proposes an improved Mamba-2-based intelligent modeling method to obtain the precise deep peak shaving process model of thermal power-molten salt heat storage coupled system. Based on the state space duality, the Mamba-2 shows great prediction accuracy and speed in dealing with the multiple long-time sequences. To efficiently model the wide-range and rapidly changing deep peak shaving sequences, the context learning capacity, computational efficiency and convergence speed of Mamba-2 must be improved. First, the sparse attention mechanism that can selectively process some of the tokens is established to promote the modeling precision of Mamba-2. Moreover, the novel least-squares interpolation (Lion) optimizer is successfully utilized to accelerate the training process in Mamba-2 for the first time. Then, both the inputs and outputs sequences inside Mamba-2 are normalized to facilitate the modeling accuracy improvement of thermal power-molten salt heat storage coupled system. The improved Mamba-2 with stable state space model, prominent modeling precision and speed effectively provides guarantee for the deep peak shaving of unit. Finally, the superior model accuracy of the proposed intelligent modeling method is verified based on the on-site deep peak shaving sequences of 660 MW thermal power unit and 40 MW molten salt heat storage system.