Abstract: With the growing demand for deep peak shaving in power systems, variable load operation of 350 MW supercritical circulating fluidized bed (CFB) boilers becomes increasingly frequent. The main steam pressure system, characterized by large time delay, strong nonlinearity and multivariable coupling, suffers from low prediction accuracy and unsatisfactory control performance. To address this issue, a deep neural network predictive control (DNNPC) strategy fused with improved Informer is proposed. IDOA-CNN-GRU-LSTM-Informer prediction model is constructed, and the improved dream optimization algorithm (IDOA) is adopted for global hyperparameter optimization to achieve high-precision dynamic and static prediction under multiple working conditions. A DNNPC system for multivariable coupling disturbances is designed with rolling optimization and feedback correction mechanisms. Simulation results show that the proposed model delivers optimal prediction performance in both dynamic and static working conditions. Compared with NNPC, PID and MPC controllers, the DNNPC controller achieves comprehensive improvement in control performance, providing theoretical and technical support for intelligent optimal control of coal-fired power units under deep peak shaving conditions.