Methanol steam reforming high temperature proton exchange membrane fuel cell (MSRFC) is a clean and efficient energy conversion device that utilizes methanol as fuel source. A MSRFC test platform with a target power of 3.5 kW was established to systematically measure and analyze the dynamic response characteristics of reformer gas composition, hightemperature fuel cell (HT-PEMFC) stack performance and temperatures of reformer, combustor and HT-PEMFC stack in the power step and stabilization periods caused by the variation of methanol fuel supply. The sample space and different machine learning methods were investigated for the accuracy and applicability of the HT-PEMFC stack based on the experimental data. The HT-PEMFC stack voltage prediction model was trained by means of Gaussian process regression. A MSRFC system simulation approach coupling the machine learning voltage prediction model and the energy conservation equations of subsystem was constructed in the frame of Simulink, which could accurately predict the power,temperature and response time of the HT-PEMFC stack during step and stabilization periods. The relative errors could be controlled within 1% and 3% respectively. The obtained experimental results and system simulation model could provide data support for optimization and scaling up of MSRFC systems.