An ICNN Embedded Day-Ahead Scheduling Strategy for Large-Scale Renewables Integration with SOFCs

Solid oxide fuel cells (SOFCs) exhibit high efficiency and fuel flexibility, making them attractive for hybrid power systems. However, their performance is highly sensitive to internal temperature and operating conditions. This paper presents a day-ahead scheduling framework that captures the nonlinear relationship among SOFC efficiency, output power, and temperature. An improved input convex neural network (ICNN) is proposed to accurately model the temperature- and power-dependent efficiency and heat generation of SOFCs. The trained ICNN is then embedded into the mixed-integer linear programming (MILP) formulation, enabling the integration of nonlinear physical characteristics into the optimization process without sacrificing traceability. To validate the proposed strategy, both offline numerical simulations and real-time experiments are conducted. The results demonstrate that the proposed ICNN-based method reduces thermal constraint violations by more than 75% compared to the Piecewise Linear (PWL) method in day-ahead scheduling, and by more than 82% compared to the constant-efficiency method. In real-time simulations, the ICNN-based strategy decreases the violation ratio from over 50% in the constant-efficiency baseline and 24–30% in the PWL baseline to below 5%, and further reduces operational costs by 13–26%.