NH ₃ /CO ₂ cascade systems have become one of the most widely used kinds of industrial and commercial refrigeration systems, in which the middle temperature and condensing temperature are the most important parameters affecting the system performance, and hence chosen as control variables in optimization control. However, the existing optimization control based on data-driven models still encounters the problem of how to obtain complete data covering a large number and wide range of operating conditions as soon and low cost as possible. To solve this problem, a novel energy-saving control method of NH ₃ /CO ₂ cascade refrigeration systems by data-driven models with varying searching boundaries was proposed innovatively. The potential of the proposed methods was evaluated from the precision and power perspectives. Results showed that the proposed energy-saving control method had the lowest RMSEs of control variables and system power, which indicated its very high optimization precision. Besides, it provided a reduction in energy consumption by 10.62% compared to the conventional constant-variable control approach and by 1.2% compared to the data-driven model with fixed search boundaries, which proved the huge potential of the proposed control method.