SOFC exhaust gas burner is an essential component of the SOFC thermal management system. The burner in the whole operating conditions of safe and stable combustion and smooth transition in the switching operating conditions is of great significance to the long-term and efficient operation of the SOFC power generation system. This paper builds an experimental test platform for the SOFC exhaust gas burner, tests the combustion performance of the burner in full operating conditions and operating conditions switching process, and establishes a prediction model for the combustion performance of SOFC exhaust gas burner by machine learning algorithm. The results show that the SOFC exhaust gas burner can meet the requirements of safe and stable combustion in all operating conditions and smooth transition between operating conditions; reliable combustion is achieved at the stable power generation condition with excess air factor of 7 ∼ 26, and the average cathode and anode pressure drop in the stable power generation condition is 142 Pa and 185 Pa, respectively; BP neural networks, support vector machines and random forests are used to establish a prediction model for the performance of the SOFC exhaust gas burner, and the random forests achieve the best prediction effect.