Abstract: In hierarchical steady-state optimization programming for large-scale industrial processes, a feasible technique is to use information of the real system so as to modify the model-based optimum. In this circumstance, a sequence of step function-type control decisions with distinct magnitudes is computed, by which the real system is stimulated consecutively. In this paper, a set of iterative learning controllers is embedded into the procedure of hierarchical steady-state optimization in decentralized mode for a class of large-scale nonlinear industrial processes. The controller for each subsystem is used to generate a sequence of upgraded control signals so as to take responsibilities of the sequential step control decisions with distinct scales. The aim of the learning control design is to consecutively refine the transient performance of the system. By means of the Hausdorff-Young inequality of convolution integral, the convergence of the updating rule is analyzed in the sense of Lebesgue-p norm. Invention of the nonlinearity and the interaction on convergence are discussed. Validity and effectiveness of the proposed control scheme are manifested by some simulations.